Consider a continuous response $Y$ and design matrix vector $\mathbf{X}$. These are related through some function $f(X) = Y$. Suppose that I am interested in estimating the probability that $Y \leq 0.1$ conditional on observing $\mathbf{X}$.

I want to use quantile regression to do this - can I confirm that this is a legitimate methodology?

We have quantiles $\tau \in [0,1]$ and after estimating our quantile regression for each $\tau$ we have our quantile estimates $\mathbf{q} := \{\hat{Q}(\tau) : \tau \in \{0.01,0.02,...,0.99\}\}$. I want to select the $\tau$ such that $\hat{Q}(\tau) \approx 0.1$. When I find such a $\hat{Q}(\tau)$ it seems to then follow naturally that $P(Y \leq 0.1) = \tau$. The reason is that my model has estimated the $\tau$-th quantile to be $0.1$, which is point on the x-axis in $Y$'s pdf that I need to find to be able to determine $P(Y \leq 0.1)$.

In practice this may not work since an estimated quantile can be lower for higher $\tau$ under some $\mathbf{X}$.

Not looking for logistic regression with a discretized response as a solution (since I already know about this).

  • $\begingroup$ A couple of clarifications: 1) Is indexing by $i$ important to the question? 2) You want $P(Y \leq y | X = x)$ but do you know the conditional distribution $P(Y=y|X = x)$? $\endgroup$ – Theja Oct 14 '13 at 16:09
  • $\begingroup$ Why not use logistic regression? You have a binary outcome: greater than 0.1 or not. $\endgroup$ – zkurtz Oct 14 '13 at 16:35
  • $\begingroup$ @Theja I'm not sure I follow. If I know the latter then I automatically know the former with integration. Also I think the notation is confused -- in the former you use $P$ to mean probability, but in the latter you use it to refer to a density function. $\endgroup$ – user2763361 Oct 14 '13 at 16:35
  • $\begingroup$ @zkurtz I had this in an old version -- looking for solutions other than logistic regression. $\endgroup$ – user2763361 Oct 14 '13 at 16:36
  • 1
    $\begingroup$ OK. Maybe it would help clarify your problem if you explain why logistic regression is not ideal here. $\endgroup$ – zkurtz Oct 14 '13 at 16:38

It doesn't appear that $Y$ is binary. Ordinal regression is a good choice here. With any of the ordinal models (proportional odds, proportional hazards, probit, etc.) you can compute the probability that $Y \geq y$ for all $y$. That probability will change at the unique values of $y$. The R rms package orm function implements this efficiently and has a function generator for exceedance probabilities. If you were extremely fortunate and really have Gaussian residuals you can use the maximum likelihood estimator of the exceedance probabilities, which is a simple function of $\hat{\mu}$ and $\hat{\sigma}$.

  • $\begingroup$ Could you please clarify "it doesn't appear that $Y$ is binary". In the question I said it was continuously normally distributed so this comment is a bit confusing to me. $\endgroup$ – user2763361 Oct 14 '13 at 16:44
  • $\begingroup$ Someone suggested using binary logistic regression but that would not be the best choice if $Y$ is not binary. $\endgroup$ – Frank Harrell Oct 14 '13 at 21:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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