Latent Causal Forest: Estimating Heterogeneous Treatment Effects under a Hidden Treatment

ICLR 2026 Conference Submission16533 Authors

Keywords: Causal inference, Hidden Treatment, Measurement Error, Causal Forest

TL;DR: We propose a forest-based framework for estimating conditional average treatment effects (CATE) under a hidden treatment.

Abstract: Learning heterogeneous treatment effects from observational data is central to applications such as precision medicine and policy evaluation. A fundamental challenge arises when treatment assignment is unobserved and must instead be inferred from noisy proxies, inducing severe measurement error. We propose the Latent Causal Forest, a forest-based framework for estimating conditional average treatment effects (CATE) under a hidden treatment. The method combines the flexibility of forest models in capturing treatment effect heterogeneity with semiparametric efficiency guarantees via efficient influence functions. We establish consistency and asymptotic normality of the estimator and develop a two-step doubly robust learner tailored to the hidden-treatment setting. Simulation studies and empirical applications demonstrate that our approach outperforms existing alternatives. To our knowledge, this is the first framework that enables CATE estimation with a hidden binary treatment.

Primary Area: causal reasoning

Submission Number: 16533