How many classes for regression problem? In my data I have a measurement which is binned into 5 classes. Those 5 bins are not only ordered, but also have the same distance between them.
My question is now if it would be better to interprete it as a normal regression problem or as a classification problem. From my intuition I want somehow use the order of the measurements for my model, but I am not sure if 5 different values are "enough" to apply regression methods. Is there any rule how many bins are required to interprete it as a regression problem?
 A: A few options:


*

*Just use linear regression.  The residuals probably won’t be normal, but worse things could happen.  On the other hand, I suspect the estimator will be biased / attenuated, maybe inconsistent.  Better minds than mine might comment on that.

*Use ordinal logistic regression.  I don’t think you can go wrong with this.  But it does throw away information—that the bins are in fact equally spaced and covering specific ranges of a numeric value.  So I suspect it’ll be inefficient.

*Use interval regression, like intreg in Stata.  The dependent variable here essentially becomes “value fell in this specific range”.  This makes better use of the data than ordinal regression.  The (small?) downside is that the assumption that the dependent variable is conditionally normally distributed can’t be strictly true—age can’t be negative, but that’s also a drawback of linear regression.

*Try them all.  If they give similar results, great!
In any case, this is not a classification problem.
A: "Normal" regression requires continuous data so you shouldn't treat it as such.
"Normal" Logistic regression requires binary variable, however there is also multinomial regression.
In the case of multinomial regression you are trying to explain a nominal variable with some regressors. You can look at the following: 
"Estimation of multinomial logit models in R: The mlogit Packages" by Yves Croissant or "mlogit" in R.
Good luck.
