Causal meta-learners (CATE estimation) in tidymodels: does anything exist, and would an extension package be welcome?

Machine Learning and Modeling
1

Hi all,

I'm a senior at Northwestern University looking for a senior thesis project to graduate
with honors in Data Science. I loved using tidymodels in my machine learning in R
courses, and I'd like to contribute something back to the community that has given me
so many resources for fun projects.

The idea: an extension package for estimating heterogeneous treatment effects (CATE) with
causal meta-learners (S-, T-, X-, R-, and DR-learners), where any tidymodels workflow can
serve as the base learner. To be clear, I mean meta-learners in the causal inference sense,
not the stacking sense used in stacks.

What I've checked so far:

  • CRAN CausalInference task view: bartCause, stochtree, hettx, FindIt, grf, subdetect,
    EffectTreat, evalITR, CRE, stepp, tools4uplift, none built on tidymodels
  • Other R packages with meta-learners: rlearner, StratifiedMedicine, and
    DeepLearningCausal, each tied to its own set of base learners
  • parsnip's grf engine: covers classification, regression, and quantile regression,
    but not causal forests
  • r-universe, GitHub, and this forum: nothing combining meta-learners with tidymodels
  • Python: CausalML and EconML offer this through a consistent API, which is roughly
    what I'd like to match

You can hand-build meta-learners from separate workflows today, but cross-fitting, uplift
metrics (Qini, AUUC), and tuning for effect estimation all have to be coded by hand.

My questions:

  1. Does anyone know of an existing package or work in progress that does this?
  2. Would there be interest in an extension package? The rough plan: learner functions that
    accept any workflow, automatic cross-fitting via rsample, tidy predict() output, and
    yardstick-style uplift metrics.
  3. For the tidymodels team: I've read the parsnip dev checklists, but a meta-learner seems
    less like a new model or engine and more like a layer over existing workflows, similar
    to stacks. Would a standalone package in that style be the right structure, or would
    you recommend something different?

A sketch of the interface I have in mind:

cate_fit <- x_learner(
  outcome_model = xgb_wflow,
  propensity_model = glm_wflow,
  treatment = "w"
) |>
  fit(data = train)

predict(cate_fit, new_data = test)

Thanks for any guidance!