linear regression

[Sonia Zam]

Hi all,

I conducted a linear regression where the dependent variable (measured as a percentage) is squared root transformed. The independent variables are categorical.

See below for the syntax

generate sqrtdepend=sqrt(dependent)
regress sqrtdepend i.indep1 i.indep2 i.indep3 i.indep4

Now I want back transform the coefficients relative to a predetermined reference level of the dependent variable so that the effect estimates can be directedly compared between variables (the reference value of the dependent variable is 30%).

The formula to do this is below.
2β*30^1/2 + β²

* β (regression coefficients) refer to differences in square-root values

Any help on how to write the syntax in stata for the above formula?

Thanks in advance for your help.

[Nick Cox]

You should cross-reference your previous thread https://www.statalist.org/forums/for...ear-regression

I guess it's no help to say that -- if I understand your goal -- I would never try to do this. Even back-transforming predictions gained on a transformed scale is messy enough. You are comparing different models; their respective output indicates relative importance of predictors in the space each is in -- with considerable caveats even then, as prediction is always a team effort.

But trying to compare coefficients estimated in different spaces looks fraught with difficulties,

It is no surprise to me that your previous version of the question didn't give you what you seek. Repeating the question seems unlikely to do so either. .

[Jeff Wooldridge]

I agree with Nick. It’s just not a good method. A better method is to divide your dependent variable by 100 and use fractional logit. See the -fragreg- command in Stata.