this is a response to a thread started a couple of months ago about

possible ways to estimate Zero-inflated Negative Binomial/Poisson

models for Panel data. I am interested in modeling differently the

zero-one distribution and the count (non-zero) distribution in my

data since 2/3 of my dependent variable's values are zero

throughout the time-span of the dataset. The count variable ranges

from 0-5.

I first followed the suggestion made in the thread to look at the

paper "From the help desk: hurdle models" by Allen McDowell,

published in The Stata Journal (2003) 3, Number 2, pp. 178–184.

What the paper illustrates is how to fit a hurdle model using ml’s

cluster(), options.

The commands are the following:

program hurdle_ll

version 8

args lnf beta1 beta2

tempvar pi lambda

quietly generate double ‘pi’ = exp(‘beta1’)

quietly generate double ‘lambda’ = exp(‘beta2’)

quietly replace ‘lnf’ = cond($ML_y1==0,-‘pi’, ///

log(1-exp(-‘pi’)) + $ML_y1*‘beta2’ - ///

log(exp(‘lambda’)-1) - lngamma($ML_y1+1))

end

You can then invoke the ml estimator with the commands:

ml model lf hurdle_ll (y = x1 x2) (x1 x2)

ml max, nolog

My question is the following: can I suggest that I am estimating or

approach an estimation of a panel data respective model if I

cluster based on each observation's identity (id) and introduce

year dummies as regressors?

Namely, the ml estimator would look like this:

xi: ml model lf hurdle_ll (y = x1 x2 i.year) (x1 x2 i.year),

cluster(id)

ml max, nolog

I look forward to receiving your insights.

Best,

Pavlos

*

* 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/