2
$\begingroup$

I have a problem where my target variable are continuous/float values in the range [0,1]. If my data were integers in {0,1} this would be a simple logistic regression / Bernoulli likelihood problem. But because my target variable are not integers, I believe that a Beta likelihood would be more appropriate.

I come from a primarily Bayesian background, so setting up a Beta likelihood in PyMC3 or Stan should be trivial. But I want to clarify if this is a sensible approach, something that's been done before, or if there is a reason that analysts don't do this.

Similarly, as my Beta likelihood model is to logistic regression, I hypothesize that a Dirichlet likelihood model would be to multi-class logistic regression. Is this sensible?

$\endgroup$
2
$\begingroup$

Yes, with continuous target you would be using beta regression and Dirichlet regression for multi-“class” problems. Usually it would be reasonable to re-parametrize the beta distribution into using mean and precision.

$\endgroup$
1
  • $\begingroup$ Oh, so these things I'm describing already exist! $\endgroup$
    – jbuddy_13
    Aug 27 at 16:15

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.