Statalist


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

Re: st: Average Partial Effects after reoprob


From   Austin Nichols <austinnichols@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Average Partial Effects after reoprob
Date   Sat, 18 Jul 2009 12:28:54 -0400

Paul Noroski<pjnoroski@gmail.com>:
The APE or marginal effect is not so straightforward to calculate for
a nonlinear random effect model, since the marginal effect depends on
the realization of the random effect, and these are typically not
estimated.  See e.g.
http://www.stata.com/statalist/archive/2009-03/msg00015.html
I don't know how -margeff- does it for -xtprobit- (but note you can
use -xtprobit- to get the APE of x on p3 as you describe it) but I
suspect it assumes the RE is zero, which can result in some very odd
results.  I think the right way to do it for a random effects model is
to simulate, i.e. draw errors from the estimated distribution over and
over, and average across results (to get SEs you could bootstrap the
simulations, to incorporate uncertainty about estimates in addition to
the distribution of REs).  That would be much more involved, of
course...

I could be wrong, of course--http://stata.com/stata11/margins.html says:
"
margins works after EVERY Stata estimation command except exact
logistic and exact Poisson; alternative-specific conditional logistic,
alternative-specific multinomial probit, and alternative-specific
rank-ordered probit; nested logit; generalized method of moments; and
structural vector autoregressive models.
"
which implies nonlinear RE models are no problem (in a week or so when
Stata 11 is out).

On Thu, Jul 16, 2009 at 2:32 PM, Paul Noroski<pjnoroski@gmail.com> wrote:
> Hello statalist,
>
> To my knowledge, the reoprob command for random effects ordered probit
> models (by Guillaume Frechette) does not lend itself to calculating
> any kind of partial effects after the regression.  In my case, I'm
> trying to get average partial effects or APE's (where the effect is
> found for each observation and the average is then taken).  I know the
> command margeff does this after some procedures, but not after
> reoprob.
>
> Does anyone know how to accomplish average partial effects after
> reoprob?  If there's no command to achieve this simply, is there a
> program I could write to get APE's (with standard errors).  For a
> model y = xb+u, suppose y takes on three values [1, 2, 3] and that we
> want the effect of x on prob(y=3).  In other words, we want the effect
> of regressor x on prob(u>cut2 - xb).  I already know how to find the
> marginal effect at the mean of other regressors, viz
>
> program define mypredict
> ...
> gen double `varlist'= 1-norm([_cut2]_b[_cons]-`xb') if `touse'
> end
>
> This gives the function that mfx will either take the derivative of
> (for continuous regressors) or take differences of (for dummy
> regressors).  Is there an analogous statement to instead get the
> average of partial effects (APE's) with standard errors after reoprob?
>
> thanks to anyone who can help,
> Paul

*
*   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   |   What's new   |   Site index