Re: st: simulating random numbers from zero inflated negative binomial estimates

["E. Paul Wileyto" <epw@mail.med.upenn.edu>]

I've never used predict with the ir option, but I assume it predicts a 
mean incidence rate GIVEN that class membership is not an inflated 
zero.  I suspect that it will not include the natural variability of the 
outcome, let alone zero-inflation.  What our simulation does is take 
those predicted linear model, adds in the natural variability for 
negative binomial, and then adds zero-inflation on top of it, all to 
reflect the natural variation you would see.
In order to gauge whether the estimate is working well, you should 
simulate the data multiple times, and generate means for the point 
estimates, and coverage probabilities.  What we did was to take our 
original model, assumed the estimated parameters are true, and then used 
them to simulate only one more data set.  Repeat that last step 200x, 
and see how often your CI includes your true value.
Looking at the script again.  This first part grabs the estimates from 
fitting your data:
zinb cignums drug  week, inf(drug  week)
predict p1 , pr
predict p2 , xb
predict lnalpha , xb eq(#3)
gen alph=exp(lnalpha)


This next part simulates the data and should be repeated many times.

gen xg=rgamma(1/alph, alph*p2)
gen pg=rpoisson(xg)
gen zi=runiform()>p1
gen newcigs=zi*pg

zinb newcigs drug  week, inf(drug  week)

There are many ways to collect the parameter estimates and CI's from the 
simulations.  I'll leave that to you.
P

On 6/3/2011 1:03 AM, Ari Samaranayaka wrote:
> Dear Paul
> Thank you very much for the great help. Your are the first person to 
> answer my question. Your answer works, and I understood the logic you 
> used in your codes. Simulated random variates goes quite closely with 
> observed data. I interpret this as a reasonable model fit. Great. 
> Thank you.
> I expected whenever the ZINB model fit is reasonably good, if we use 
> the ZINB postestimation predict command to produce predicted numbers, 
> those predicted numbers also should goes closely with observed data.
> For example, if I use the command
> predict expec, ir
> then distribution of resultant values in "expec" should have similar 
> distribution to observed data (because we do not specify an "exposure" 
> in our model). However those 2 distributions quite different. Did I 
> misinterpret the result from predict command.
> Thank you again
> Ari
--
E. Paul Wileyto, Ph.D.
Senior Research Investigator, Department of Biostatistics&  Epidemiology
Director of Biostatistics, Tobacco Use Research Center
School of Medicine, U. of Pennsylvania
3535 Market Street, Suite 4100
Philadelphia, PA  19104-3309

215-746-7147
Fax: 215-746-7140
epw@upenn.edu

*
*   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/