Regression with skewed data

noetsi

Fortran must die
#1
Our data is highly skewed, so much so that we routinely report median values not average ones in descriptive models. Now I am running regression with an interval variable I know will be skewed. I was wondering [given that I prefer not to transform variables if possible] if there was a regression that used medians rather than averages and what the good and bad points of doing this are?

I know there are regression that work with medians, but I have never worked in this area.
 

CowboyBear

Super Moderator
#2
Great that you're cautious about transformation. But haven't we talked a fair bit here about the fact that regression does not assume a normally distributed DV? Are your errors highly skewed? Is your sample very small?

That aside, I think the technique you're thinking of might be quantile regression?
 

vinux

Dark Knight
#4
Our data is highly skewed, so much so that we routinely report median values not average ones in descriptive models. Now I am running regression with an interval variable I know will be skewed. I was wondering [given that I prefer not to transform variables if possible] if there was a regression that used medians rather than averages and what the good and bad points of doing this are?

I know there are regression that work with medians, but I have never worked in this area.
You need to see how the residual is distributed.