HTERF: Transfer learning boosts causal inference with CATE bounds

A new preprint (arXiv:2606.07693) from Bérénice-Alexia Jocteur, Véronique Maume-Deschamps, and Pierre Ribereau tackles the challenge of applying causal forests when target domain data is scarce. Their method, HTERF, extends transfer learning to estimate Conditional Average Treatment Effects (CATE) — the expected impact of a treatment given specific covariates. The key innovation is adapting the offset method (Wang, 2016) to a causal context: intermediate models capture the shift between source and target distributions, allowing the causal forest to adjust its predictions.

The authors prove a bound on the CATE error in the target domain, expressed in terms of the error of those intermediate models. Simulation studies across diverse settings and a real-world dataset demonstrate that HTERF significantly outperforms naive applications of causal forests on small target samples. This work opens practical pathways for deploying causal inference in fields like medicine or economics where gold-standard randomized trials are expensive but transferable knowledge from related populations exists.