st: Modeling Options Overdispersion and "excess zeros"

[celdjt@umich.edu]

I am working with data that gives clear evidence of both overdispersion 
and excess zeros. Accordingly, I have been examining the viability of 
modelling it using zero-inflated models (ZIP and ZINB) as well as 
negative binomial regression (NBRM). Vuong tests tell me that ZIP is 
preferred to the standard Poisson; on the other hand, ZINB does not 
give a significant improvement over NBRM, implying (as I understand it) 
that either a zero-inflating process or between-subject heterogeneity 
may account for both the overdispersion and the excess zeros in the raw 
data. At the same time (and similar to the the example given in David 
Dukker's reponse to an issue like this in FAQ), standard errors for 
ZINB are large, suggesting a poor fit of this model. Subsequently, I 
followed that procedures given in Long and Freese (2e: 2003, pp. 
283-84) and graphed the differences between the observed probabilities 
and mean predictions for the different models. The results indicate 
that NBRM performing more poorly than ZIP, with ZIP and ZINB performing 
in an almost identical manner. Subsequently (and again following Long 
and Freese) I have compared the models using BIC', with ZIP emerging as 
overall best. From this, I have concluded that ZIP is my best option. 
At the same time, it also fits with where I am at substantively with 
these data. To those who actually have knowledge of the underlying 
processes associate with this, does all this seem reasonable? Have I 
left something out or made a leap I shouldn't have? Thanks.
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/