What nonlinear model to use?

[Angelo Nunnari]

Hi everyone!

I am doing a research paper for the university. I am using some data coming from the European Social Survey, the first problem I met was the impossibility of using the design weight, the post-stratification and population correction to have good data because of too many variables specified, i removed nearly the 80% of them, but it is not still working.

Bypassing this question, i am doing a regression to explicit the determinants of a count variable which is drawn on a 0-10 scale. What is the non linear model to use? I know that in this case i should use the poisson one. However I don't like it very much, it is good to use a logit or probit even though they are better for 0 1 dichotomous variables?
Are there any other useful models to use?

Finally, i would like to compare the marginal effects obtained by two different regression, should I usde the mean comparison tool?

Thank you very much for your help,

best regards

[Marcos Almeida]

Welcome to the Stata Forum / Statalist,


With regards to several parcels of your query, I suggest you try to be more specific. For example, instead of "it is still not working", you could present command and output as recommended per FAQ.

I also failed to understand how the count variable has a 0-10 scale, what you meant by "good data" and, last but not least, the removal of 80% "of them" (observations? variables?).

Now, going directly to model selection, however much you dislike it, since you have count data, you're supposed to think about selecting a Poisson model or, in case of, overdispersion, a negative binomial regression.

[John Mullahy]

Angelo: Given the doubly bounded nature of your 0-10 outcome variable, perhaps also consider

1. glm, with a binomial family specification (and maybe a logit or probit link)

and/or

2. fracreg

Good luck, John