In this study, we address the challenge of using machine learning to predict individual treatment effects (ITEs). Previous research has focused on developing machine learning models that can estimate the conditional average treatment effect (CATE). These models combine intermediate estimates to produce point estimates of CATE.
In this paper, we propose a new approach called conformal meta-learners. This framework uses the standard conformal prediction procedure on top of CATE meta-learners to provide predictive intervals for ITEs. We specifically focus on a class of meta-learners based on two-stage pseudo-outcome regression and develop a stochastic ordering framework to assess their validity.
Our analysis shows that conformal meta-learners yield marginally valid inference if their conformity scores, based on pseudo outcomes, stochastically dominate conformity scores evaluated on unobserved ITEs. We also demonstrate that commonly used CATE meta-learners, such as the doubly-robust learner, satisfy a model- and distribution-free stochastic dominance condition, enabling valid conformal inferences for practical levels of target coverage.
Unlike existing methods that conduct inference on nuisance parameters, our conformal meta-learners enable direct inference on the target parameter, ITE. Through numerical experiments, we show that our approach provides valid intervals with competitive efficiency, while maintaining the favorable point estimation properties of CATE meta-learners.