Betareg

[Sonia Zam]

Hi all,

I have a propotion dependent variable that’s greater than zero and less then one. The mean is 0.1240891 and SD 0.1363 and its positivly skewed. All the independent variables are dummy variables. I have read that betareg is most appropriate model for propotion data.
Want to check did I use the right model and also do I need to check any assumptions before carrying betareg ?

Betareg dep i.var1 i.var2 i.var3 i.var4
margins dep

After running estat ic to check the model fit AIC and BIC are approx -4400

Thank you in advance.

[Maarten Buis]

It is important that you tell us exactly what you did and exactly what you want to achieve. Potential problems are almost always somewhere in the details. So if you don't give us the details, we obviously cannot help you. In your case the first line should have returned an error message as there is no such command Betareg, however betareg certainly exists. The second line should also have returned an error message. What you are probably looking for is margins, dydx(*). However, I am only guessing what you want, so there is no guarantee that this is actually what you want. Without you telling us what you want to achieve, there are infinite possibilities of miscommunication.

One of the properties of a beta regression is that the conditional mean also plays a role in the conditional variance ( the variance is smaller the closer the mean gets to either the lower or the upper bound). This is a reasonable model for the heteroscedasticity. However, if there is additional heteroscedasticity on top of the "normal" heteroscedasticity in a beta distribution, then the parameters for the mean function try to do two things at the same time: best fit the conditional mean and best fit the variance. As the two no longer fit together you will get biased estimates. One way around that is to more completely model the heteroscadasticity by adding your explanatory variables also to the scale() option in betareg. This way the parameters in the mean function no longer have two competing objectives. Alternatively you can use a fractional logit (fracreg) , which only focuses on the conditional mean. Personally, I would stick to a fractional logit if I was only interested in effects on the conditional mean, and move to beta regression if I was interested in other facets of the distribution, like the conditional variance.