We measured drinking frequency among early adolescents using an item with response options: never, 1-4 times in life, 3-4 times in last year, once per month or more, once per week or more, every day. Responses are positively skewed. We modeled those as count variables and did some path modeling using the negative binomial distribution. Of course the values are not literal counts, but it seemed like a good way to capture what we had. Is there a better alternative for modeling ordered categorical variables with the positive skew? If so, what would it be? Thank you.
I would not treat this as a count variable. I would treat it as an ordered categorical variable. Categorical data methodology can handle floor and ceiling effects. Skewness is an issue for continuous variables.