Ordinal outcome support

[ck37]

Hello,

Thank you for an excellent package and method. Is there any potential interest in supporting ordinal outcomes with causal forest? I couldn't find anything in the docs or issues thus far, and would be happy contribute a little time to make that happen if it's of interest.

I did find this ordinal version of random forest that could be helpful to review:
https://link.springer.com/article/10.1007/s00357-018-9302-x

Appreciate it,
Chris

[erikcs]

Hi @ck37, thank you for the suggestion, it sounds interesting! Do you have a particular application in mind, and what causal estimand would you envision targeting?

A related question is whether there exists an orthogonal estimating equation in this setting that could be paired naturally with the GRF/causal forest machinery

[ck37]

Hi @erikcs,

Thanks for considering it. For me the most common ordinal outcomes are: 1) rating scales (e.g. clinical judgment of suicide risk: minimal, low, moderate, or high), and 2) disease severity grading (e.g. mild, moderate, severe).

From psychometric theory single item ratings are fundamentally ordinal, and a binary outcome is basically a special case of an ordinal variable. That said, where possible we would prefer to combine multiple items via item response theory into an interval-valued, continuous measure.

I am not sufficiently theoretically skilled to be certain about orthogonal estimating equating options but my sense is that following something from the work of @idiazst would be informative on that and the causal estimand:

• 2016: https://onlinelibrary.wiley.com/doi/abs/10.1111/biom.12450
• 2021: https://academic.oup.com/biometrics/article/77/4/1467/7446192
• 2022: https://academic.oup.com/jrsssa/article/185/4/2156/7069383

Happy to help out an dig further into this though if it would be useful.

Cheers,
Chris

[erikcs]

Thanks, @ck37! It sounds like this could take the form of research project, I am happy to help if you wish to take the lead. You can essentially think of causal forest as a fast way to fit a linear regression, but conditional on X. This brief overview might be helpful. What you'd want to start with is an estimating equation that delivers the estimand you care about, ignoring covariates. You can then pair this with grf forest weights to target this estimand conditional on X (e.g, the CATE). Further down the line, to organize this, it probably makes sense to start a new repo with a stripped-down version of grf that can be used to build a causal forest for ordinal outcomes only. I am happy to help with that!

[ck37]

Ok, I will see if I can either start an ordinal branch or a stripped-down ordinal grf one of these months. At least for my own reference, it may also help to review @benkeser's https://github.com/benkeser/drord package.