Re: st: hurdle nb model marginal effects?

["Dimitriy V. Masterov" <dvmaster@gmail.com>]

Brent,

I am not sure if I have a good answer for the first two questions
about the derivative of the product and its variance. Winkelman's
count data book punts on this question.

The quantity you want might be more easily obtained from zero-inflated
count data model (user-written zip or zinb) with predict and margins,
but I don't know if that model is appropriate for your data.

The third one is easier. If you use the factor variable operator (like
i.dummyvar) in your specifications, margins will use the finite
difference method to figure out the MEs. I am not sure how
ztnb/zip/zinb handle factor variable notation. Also, forgot that ztnb
has been superseded by tnbreg, which handles the factor variables
nicely (like probit and logit).


DVM

On Mon, Mar 18, 2013 at 8:30 AM, Brent Gibbons <brent.gibbons@gmail.com> wrote:
> Dimitriy, thanks for reminding me about the fact that the hurdle-nb
> model can be decomposed into a logit and a zero-truncated nb. To
> explain my concerns, let me label the binary outcome of the logit by
> B=1 if in fact the response is non-zero (and B=0 otherwise) and let me
> label the continuous outcome of the truncated nb as Y.
>
> The combined marginal effect of a continuous X variable on the
> conditional expected value of the outcome is = d[E{(B|X)*(Y|X)}/dX.
> (Consider "d" as the relevant symbol for the partial derivative. Sorry
> I can't do the correct symbol in g-mail.)
>
> The first problem is how do I compute this derivative if all I have
> are d[E(B|X)]/dx and d[E(Y|X)]/dx?  Can independence of the
> conditional expectations be assumed?
>
> I assume that if this derivative can be computed, I would follow the
> procedure used in the Stata "margins" command and just take the sample
> average of this derivative. but this leads to the second question: how
> should I compute the estimated variance of this sample average of
> d[E{(B|X)*(Y|X)}/dX?
>
> The third question arises when the X in question is a dummy variable
> rather than continuous. How do I deal with the computation of the
> average differential (rather than derivative) in dealing with marginal
> effects of dummy variables?
>
> Thanks for any suggestions.
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/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/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/