Regression for dependent variable that is categorical

[Rochelle Zhang]

Dear Stata listers,


My dataset is panel data. My regression dependent variable is non negative integer (a count). would it be correct to think my choice of regression models are between Poisson or binomial. but not traditional reg?

thanks,
Rochelle

[Clyde Schechter]

Regression models for count variables often use the Poisson or negative binomial models. But ordinary linear regression is not necessarily out of the question. When the expected values are large, the Poisson distribution looks a lot like a slightly discretized version of the normal distribution. There is also the question of whether a logarithmic link is a good reflection of how your dependent variable is related to the linear predictor. (In fact, your question can be stood on its head. It is often the case that with a continuous variable, a better approach than regressing the log of that variable against linear predictors is to use Poisson regression. http://blog.stata.com/2011/08/22/use...tell-a-friend/.)

So I think you need to do some exploratory data analysis first to get a sense for which type of model is the best specification of the relationships in your data. You might also need to give some deference to whatever traditions prevail in your discipline about these matters.

[Maarten Buis]

For panel data you need to look at an xt command. For count data poisson is a good starting point. Combining the two suggest looking at help xtpoisson. For more flexibility you can look at help mepoisson or help menbreg.

[Rochelle Zhang]

Thanks very much Maarten for suggesting the specific commands !!!

Thanks very much Schechter for your detailed response!!!! I agree that I should explore the data to see which distribution fits my data better as well as following the prior work in my area.

May I ask how we explore the distribution of my data , e.g. the dependent variable , which command ? do I do a plot ?

[Nick Cox]

You could always look at the distribution with a histogram. However, it is not necessary for the marginal distribution to be Poisson (or whatever) as the model is at most about its distribution conditional on covariates. Moreover, although the names are evocative, they are in many ways misplaced. The more important fact about Poisson regression is its use of a logarithmic link. In fact, Poisson regression will often give a sensible model even with continuous responses, although you need to use robust standard errors. Loglinear regression would be a better name, and it's in some literature.

[Carlo Lazzaro]

As an aside to previous sound advices, Rachel may want to take a look at Joseph Hilbe's Negative Binomial Regression. 2nd edition. Cambridge, 2011, (http://www.stata.com/bookstore/negat...ion/index.html), especially chapters 7-10 and 14.

Kind regards,
Carlo