Linear models are models like $\mathbb E\left[Y\vert X\right]=X\beta$.

Linear quantile regression models replace $\mathbb E\left[Y\vert X\right]$ with $Q_{\tau}\left(Y\vert X\right)$ for conditional quantile $\tau$, giving $Q_{\tau}\left(Y\vert X\right)=X\beta$.

Generalized linear models include a link function to transform the conditional expected value of a linear model, so $g\left(\mathbb E\left[Y\vert X\right]\right)=X\beta$.

Is there any role for generalized quantile regression that transforms the conditional quantile $\tau$, like $g\left(Q_{\tau}\left(Y\vert X\right)\right)=X\beta$? How would we go about developing a likelihood function to maximize or a loss function to minimize?


1 Answer 1


Bounded distributions.

We have linear probability model that comes with a drawback of producing fitted probabilities that are outside of the [0,1] interval for some values of the regressor $X$. We have a generalized linear model – logistic regression – that fixes that.

Suppose we want to model the quantiles of a bounded distribution such as $\text{Uniform}(0,1)$. By using a linear model, we would face the same type of problem as described above. It would be nice to have the fitted quantiles inside of the [0,1] interval. This is where generalized quantile regression of the logistic type (but not only) would come in handy.

I have no thoughts about maximum likelihood estimation as of now.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.