Which regression approach can I use?

[Paul Olz]

I am looking at data about various employees in the US. Per employee, I have the age (metric), the degree of education (ordinal scaled) and the subject of the studies he/she did (nominal scaled). I want to look at the effect that those three variables have on a metric characteristic ratio (0 low to 1 complete). Unfortunately, I do not get information about which regression type I need to use. Any information about the approach would be very helpful.

[Rich Goldstein]

h fraclogit

[Paul Olz]

Thank you very much. Unfortunatly, fractional regressions are only useful for binary variables. My dependent variable is more or less a percentage ranging from 0 to 1.

Therefore, I think the fractional logic regression is not well suited, right?

[Carlo Lazzaro]

Paul:
if your regressand is continuous on the 0-1 scale and there's no censoring, you may give -regress- a shot.

[George Ford]

Unless the mean of the dependent variable is near 0 or 1, regress is a decent place to start. I wouldn't predict from regress, however, as it may produce values outside the unit interval.

Two comparable options that match your needs are:

fracreg

glm y x1 x2 x3 , family(binomial) link(logit) vce(robust)

[Carlo Lazzaro]

Paul:
George's reply is wiser than mine.

[Rich Goldstein]

re: #3 - no that is not right; also note that #5 is essentially the same answer as #2

[George Ford]

Yes. fraclogit is available as well. It is not a 0/1 only model, but a unit interval model.

[Paul Olz]

Thank you very much, George! According to the descriptions and the additional paper I looked at, this definetly fits my needs. Which numbers do tell me in the output now if the usage is valid or not? I can not find any interpretation guidance. Can you help?

[George Ford]

If you've got a unit interval dep variable and use a unit interval model, then the model is correct in that regard. Specification error and so forth is a different matter. -fracreg- and -glm- should provide nearly identical results (behind the scenes I believe them to be nearly the same).

Given that the coefficients are not directly interpretable, I'd use -margins-post estimation to get something useful.

https://www.stata.com/features/overview/fractional-outcome-models/

[George Ford]

https://www.youtube.com/watch?v=mJzrWocdWGY