Problem with Count data model

[JP Doorn]

Dear all,

Currently I'm doing a reseach on a dependent variable that is a (non-negative) integer count variable, in a panel dataset with quarterly time units. The panel variable (countryid) can take 6 different values and I study 70 quarters of data. Furthermore, I will test the data in 5 subdivisions of industries (total/all industries, non-manufacturing, manufacturing and high/low R&D industries). My model contains 4 independent variables, time trend and year dummies (to capture year specific effects) I am not an experienced researcher and just a Stata layman. I have a few questions and dearly hope that you can help me with my problems.

I substantially prefer to use the OLS (xtreg) model (adjusted with fixed or random effects) because of my familiarity with it. Is it possible to perform sufficient tests with OLS? Do you have any recommended papers/books that support the use of OLS in this case?

The paper where I got the initial idea from for my research, uses a random effects negative binomial model, but does not elaborate how he uses it. This paper also does not look at the relationship of multpile countries (my panel variables) and only looks at one country at another country. The model should give better estimates than the Poisson model, because it allows for overdispersion. (in my case, mean<sd of the dependent variable)
Reading a lot of papers and files on this model, does still not give me a clear summary image of this model. What does the output mean (the itterations on top, the log likelihoods etc.) And do I need to check for heteroskedasticity, serial correlation and what other things, and how do i correct this model for these things? The last question, when I try to perform a Hausman test to see if xtnbreg, fe or xtnbreg re is the best model to estimate the relationships, the Hausman test gives a negative test statistic and no result, how do i test which model is the preferred one?

Thank you so much for your time.

Kind regards.

[Phil Bromiley]

You may not have received an answer because you're not asking a Stata question per se, and you've actually asked a large number of questions. See the FAQ on asking questions. This list is better for questions where you can show us what you ran, what Stata gave you, and what you wanted (e.g., in data manipulation or model to be estimated).

First, with a count dv, you need a technique designed for counts - that means xtpoisson or xtnbreg or maybe a zero inflated estimator (if you have an excess of zeros). While some use regression in such cases, the count models are the correct way to go.

You need to read econometrics texts and the Stata manual on count models in panel data. Once you've gotten a bit clearer on what you don't know, you might want to pose a clearer question to the group.

[Carlo Lazzaro]

JP (the preference for real full (given) names on this forum still applies. Please see the FAQ on that topic. Thanks):
welcome to the list.
As an aside to Phil's helpful comments, the following Joe Hilbe's valuable textbooks can help you out in dealing with count data regression:
http://www.stata.com/bookstore/negat...ion/index.html
http://www.stata.com/bookstore/modeling-count-data/