Re: st: ZOIB procedure

[Prerna S <maruiprerna@gmail.com>]

Thanks you Maarten. I'm new to Stata and the list so I apologize for
breaks in protocol.

I did acquire zoib through ssc.

I am much more concerned about the interpretability of the
coefficients so perhaps Atchison (2003)'s method may not be suitable.
I'll take a look at it nevertheless. With respect to dirifit, I
interpreted it as being similar to betafit in that it does not handle
'0' and '1' very well. Both dirifit and fmlogit, moreover, do not
handle the sample selection problem like zoib. Is that a correct
inference?

With respect to endogeneity, I was wondering if it would be suitable
to apply the SMith-Blundell procedure described by Wooldridge (2002)
Econometric Analysis of Cross Section and Panel Data, MIT Press.

Thanks for the Hardin and Hilbe reference

Prerna Marui S.

On 19 September 2011 04:11, Maarten Buis <maartenlbuis@gmail.com> wrote:
> On Sun, Sep 18, 2011 at 11:51 PM, Prerna S <maruiprerna@gmail.com> wrote:
>> I am using zoib (Stata 11.2) to estimate 2 proportions. I have 3 questions.
>
> -zoib- is a user written program. The Statalist FAQ asks you to
> specify where you got it from. This is not to annoy you, but to help
> you. Experience learns us that there are often different versions of
> user written software floating around in cyber space. How can we help
> you if we do not know which version you are talking about? I am
> assuming you are talking about -zoib- as available from SSC.
>
>> 1. The model that I am attempting to estimate looks like the following.
>>
>> p1 = a1 + a2X + e1
>> p2 = b1 + b2X + e2
>>
>> where pi is the proportion of income from source i.
>>
>> Is there any procedure that approximates an SUR for zoib that I can
>> use? I tried the suest option but it does not offer a test statistic
>> and the results under suest appear to be the same as without suest. I
>> am unsure if this implies that SUR does not matter or if I missed
>> something.
>
> -suest- only changes the variance covariance matrix, such that you can
> perform tests across models. Like the -vce(robust)- option, they do
> not change the estimated coefficients. To perform those tests, you
> need specify them yourself using -test-. For more see the manual
> entries of -suest- and -test-.
>
> I am guessing, but it seems to me you are worried about correlation of
> error terms across income sources. This is a hard problem, in part
> because proportions are inherently (negatively) correlated. If one
> proportion increases, than the rest will have to decrease. Some work
> has been done by Aitchison (2003), but he sacrifices an interpretable
> effect of explanatory variables on the proportion in order to get the
> correlations right. This is fine if you are mainly interested in those
> correlations, but a problem otherwise.
>
> Alternative solutions are -dirifit- and -fmlogit-. The former makes
> the strong assumption that the correlations between proportions are
> only due to that automatic correlation that occurs between
> proportions. -fmlogit- uses a quasi-likelihood argument to by-pass
> that entire problem. Both -dirifit- and -fmlogit- can be downloaded
> from SSC. They are discussed in this talk at the 2010 German Stata
> Users' meeting: <http://ideas.repec.org/p/boc/dsug10/04.html>.
>
>> 2.  Wooldridge (2002) suggests using Smith-Blundell/Rivers Vuong
>> method that for dealing with endogeneity with respect to fractional
>> logit and tobit. Is this for some reason unsuitable for zoib?
>
> The Statalist FAQ asks you to provide full references rather than a
> name/year reference.
>
>> 3.  I would like to investigate the usual problems like
>> multicollinearity, heteroscedasticity, non-normality. Is there a
>> resource that I might refer to for regression diagnostics for zoib?
>
> The usual problems are only usual in the linear regression case. You
> will need to redefine them when dealing with other types of models.
> Multicolinearity is never a problem, it is just an unfortunate
> historical accident that it ended up in that row of "problems".
> Heteroskedasticity, is in and of itself not a problem when dealing
> with a (zero-one-inflated) beta distribution, as that distribution
> allows for heteroskadasticity. This leaves the question open whether
> it does so correctly. Normality is only a problem when you assume a
> normal distribution, which is obviously not the case when you estimate
> a zero-one-inflated beta distribution. This leaves the question
> whether your data is well enough approximated with a zero-one-inflated
> beta distribution. In short you need to rethink what those "problems"
> mean in the context of your model. A good source is chapter 4 of
> Hardin and Hilbe (2007).
>
> Hope this helps,
> Maarten (author of -zoib-)
>
> Aitchison (2003) The Statistical Analysis of Compositional Data.
> Caldwell, NJ: The Blackburn Press.
>
> James W. Hardin and Joshep M. Hilbe (2007) "Generalized Linear Models
> and Extensions", second edition. College Station, TX: Stata Press.
> <http://www.stata.com/bookstore/glmext.html>
>
> --------------------------
> Maarten L. Buis
> Institut fuer Soziologie
> Universitaet Tuebingen
> Wilhelmstrasse 36
> 72074 Tuebingen
> Germany
>
>
> http://www.maartenbuis.nl
> --------------------------
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/
>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/