Research map / Probabilistic and scientific modeling / Treatment effects

Doubly robust identification of treatment effects from multiple environments

ICLR 2025 · Probabilistic and scientific modeling › Treatment effects · approach: Latent-variable causal inference · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RAMEN, a multi-environment method that identifies the average treatment effect under bad controls and unobserved variables by minimizing a kernelized invariance loss over covariate subsets, with doubly robust identification via either treatment or outcome invariance.

Problem

Estimating average treatment effects from observational data when the causal graph is unknown and the covariates may include bad controls (post-treatment variables) and unobserved confounders.

Approach

Find a covariate subset whose conditional mean of the treatment or outcome is invariant across environments, using a kernelized invariance loss; the adjustment set is the one with minimal loss, and identification holds if either T or Y has its parents observed and is invariant (doubly robust identification).

Key results

• On known DAGs with n=2500 and 5 environments, RAMEN variants achieve consistently lower ATE MAE than IRM, adjust-all, and null baselines across invariance scenarios.
• On 100 random Erdos-Renyi DAGs (p=20), the Gumbel variant significantly outperforms all baselines across numbers of environments, while IRM fails to beat trivial baselines even with many environments.
• On the maternal smoking data, the Gumbel variant estimates a birth-weight reduction of -214.60 g (std 25.20), within the 200-250 g range reported in epidemiological literature.

Measured effects reported in the paper

Surprising findings

• Adjusting for the null set performs competitively on IHDP because confounders have limited impact on outcome and treatment there.
• IRM-based treatment effect estimation performs worse than adjusting for all covariates when the treatment is not invariant.

Limitations

• Kernel-based loss suffers from curse of dimensionality and quadratic computational cost.
• Requires sufficient heterogeneity across environments, which may be too strict in practice.
• Combinatorial subset search is expensive and the Gumbel trick is a heuristic; performance degrades when no invariant node is present.

Experimental setup

Datasets: IHDP maternal smoking birth weight dataset (Cattaneo)

Benchmarks: synthetic known DAGs random Erdos-Renyi DAGs IHDP semi-synthetic maternal smoking real-world

Metrics: mean absolute error of ATE across environments

Baselines: IRM for treatment effect estimation adjust for all covariates no adjustment (null)

Builds on: invariant risk minimization backdoor criterion invariant causal prediction Gumbel-softmax

Closest papers by meaning

• Benchmarking Estimators for Natural Experiments: A Novel Dataset and a Doubly Robust Algorithm (NeurIPS 2024 · similarity 0.89)
• Learning Representations of Instruments for Partial Identification of Treatment Effects (ICML 2025 · similarity 0.88)
• Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.88)
• Continuous Treatment Effects with Surrogate Outcomes (ICML 2024 · similarity 0.88)
• Meta-Learners for Partially-Identified Treatment Effects Across Multiple Environments (ICML 2024 · similarity 0.88)
• Coupling Generative Modeling and an Autoencoder with the Causal Bridge (NeurIPS 2025 · similarity 0.88)