Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.


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

Re: st: ivpois with a binary endogenous predictor


From   Austin Nichols <austinnichols@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: ivpois with a binary endogenous predictor
Date   Sat, 22 Sep 2012 16:04:03 -0400

Dimitriy V. Masterov <dvmaster@gmail.com>:
First, the blog post you reference ends with a reference to the
presentation that inspired it:
http://www.stata.com/meeting/boston10/boston10_nichols.pdf
so you may want to start there.  When you regress ln(y) on X, and you
observe cases where y==0 or y<0, not only are you using the absurd
model that E[ln(y)|X] is linear (impossible since ln y is undefined in
sample, so the mean is undefined) instead of assuming that
ln(E[(y)|X]) is linear (not impossible), but you are dropping cases
with y<=0, and selecting on the dependent variable is not going to
give unbiased estimates, no matter how good your instrument is!
Finally, the coefs of -2.4 and -4.7 translate into a 90% and 99%
reduction, respectively.  Always think about what the corresponding
positive coef would mean when translating these percentages in your
head: a 2.4 coef implies levels 11 times higher, or 1000 percent
higher. If a coef of 2.4 implies outcomes 11 times higher, a coef of
-2.4 has to imply outcomes 1/11 as high.

On Fri, Sep 21, 2012 at 9:26 PM, Dimitriy V. Masterov
<dvmaster@gmail.com> wrote:
> Thanks for the quick reply, Austin. I have 3 follow-up questions. I
> tried the following command:
>
> ivpois y a b c, endog(x) exog(z)
>
> The coefficient on the binary endogenous variable x is -2.4. That's
> the expected sign (compared to non-IV poisson), but the magnitude
> seems too small. Using the dummy elasticity calculation for x going
> from 0 to 1, -2.4 translates into an (exp(-2.4)-1) = 0.9% reduction in
> y. Is that interpretation correct?
>
> I also tried this specification with -ivreg2- and ln(y), which gives
> me a coefficient of -4.7, which is almost -1%. Are there any other
> methods I can shoehorn this into to check my estimate?
>
> In his Stata Blog post*, William Gould wrote that that Poisson
> regression with the Huber/White/Sandwich linearized estimator of
> variance is a permissible alternative to log linear regression. Can
> ivpois incorporate this? It does not let you specify vce(robust).
>
> * http://blog.stata.com/2011/08/22/use-poisson-rather-than-regress-tell-a-friend/
*
*   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/


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