Re: st: suggested references about the variables to include in zero-inflated portion of zinb?

[Steven Samuels <sjhsamuels@earthlink.net>]

Tim--the Subject of your last post was completely uninformative (st:  
Re: statalist-digest V4 #3224). If you receive the Digest, do not use  
the "Reply" button to respond.


I have a few thoughts:

1. The reviewer's original opinion is not correct. If your target  
parameter is the mean score, then OLS may give a consistent estimate,  
even if the data are skew and non-normal. The proviso is that you  
have a good prediction model for the mean.  However with OLS,  
standard errors will be incorrect. The fix is easy: -reg- with a - 
robust- option will give standard errors that are model-free.

2. Did you compare observed and expected values by eye and with a chi  
square test?  If the -zinb- fit is not good, there is little  
justification for using it.

3. If, by chance, -zinb- happens to give a good fit, standard errors  
based on the ZINB model will be wrong. You should use the  -robust-  
option or a bootstrap, as Carlo suggested.

4. Published analyses of CESD with the zero-inflated negative  
binomial are not, in themselves, justification for using -zinb- in  
your problem. Did the published distributions fit the data?  I've  
done analyses with full and reduced versions CESD. In one data set  
and in national data the distribution was quite symmetric. In another  
data set the distribution was bimodal. (I think this was an  
interviewer problem) In neither case was there a lump at the minimum  
(or maximum) value.  In fact, the extreme responses were the rarest  
ones.

5. If you do see lumps at the extremes, considered that they are  
dishonest. Why? With count data, a separate model for responding at  
all is plausible.  With questionnaire scales, a minimum or maximum  
score is the result of a respondent checking the same value for  
every  item. (I use the world "lumps", but in the statistical  
literature, isolated higher density regions are usually called "bumps".)

6. If you want to fit the distribution of scores, as opposed to  
predicting means, the beta distribution may provide a good  
approximation. Divide the scores by the maximum possible, so that the  
results are proportions. Then download -betafit- from SSC. You will  
need to add a small constant to the zeros and subtract it from the  
ones before you do your regressions.




-Steve

> I am using zinb to estimate level of psychological distress (scores  
> range from 0-24) using various demographic variables and measures  
> of use of the Internet. I've used -countfit- to compare various  
> count models and the results support zinb as the best fitting model.
>
> I am uncertain, however, about how to justify the variables that I  
> include in the zero-inflated part of the model. I've read journal  
> articles that have used zinb, read the book by Freese and Long, and  
> searched the Internet and Statalist but I have not been able to  
> find any detailed recommendations or procedures. Can anyone suggest  
> any other sources (books or journals) that provide an explanation  
> or a good example of this process?
>
> Ideally I would like to find a good source that I can cite in the  
> paper -- but I appreciate any suggestions about this you might have.
>
> Thanks for you help,
> Tim
>
> -----------------------------------------------------
> Timothy M. Hale, MA
> Graduate Assistant
> University of Alabama at Birmingham
> Department of Sociology
> email:  timhale@uab.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/

*
*   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/