Research map / Probabilistic and scientific modeling / Treatment effects

Estimation of Treatment Effects in Extreme and Unobserved Data

NeurIPS 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 the Normalized Extreme Treatment Effect (NETE), a causal estimand for treatment effects on rare extreme events modeled by multivariate regular variation, with IPW and DR estimators and non-asymptotic error bounds.

Problem

Estimating causal treatment effects on rare, extreme outcomes (e.g., extreme climate events) from observational data, where extreme samples are scarce and extrapolation is needed.

Approach

Combines causal inference (exogeneity, overlap, double machine learning) with extreme value theory: under multivariate regular variation and asymptotic homogeneity, the tail treatment effect factorizes into a spectral-measure average of the conditional effect times an EVI-based tail moment, which is estimated by IPW/DR plus an adaptive Hill estimator.

Key results

• On synthetic data, EVT-DR and EVT-IPW generally achieve lower MSE than Naive-IPW and Naive-DR across configurations of alpha, beta, d_u, d_z (averages over 50 runs).
• On the semi-synthetic wavesurge setup, EVT-DR deviates by at most 0.3 from the test-set NETE surrogate, while naive estimators overshoot by about an order of magnitude (e.g., Naive-DR 41.93 vs. 0.13 for (2,2)).
• Under the stated rates, the DR estimator's error bound matches the rate of prior work without prior knowledge of the tail index beta, given negligible e(t) and known alpha.

Measured effects reported in the paper

Surprising findings

• MSE sometimes increases with more samples in the synthetic experiments, which the authors attribute to violation of the Pareto-type assumption making the threshold rule inapplicable.
• The naive estimators overshoot the surrogate NETE by roughly an order of magnitude on the semi-synthetic data.

Limitations

• The scaling exponential alpha is estimated heuristically via log-log regression without theoretical guarantees.
• Non-asymptotic analysis relies on a restrictive Pareto-type tail model (Assumption on linear transformations of approximate Pareto variables); the authors leave generalization to future work.
• The threshold selection rule may fail when the tail model is violated, leading to variance-dominated error.

Experimental setup

Datasets: wavesurge dataset (wave and surge heights off south-west England) synthetic Pareto-generated extreme noise

Benchmarks: synthetic simulations (MSE vs. known NETE) wavesurge semi-synthetic dataset

Metrics: mean squared error estimated NETE relative to test-set estimate

Baselines: Naive-IPW Naive-DR

Builds on: double machine learning adaptive Hill estimator inverse propensity weighting doubly robust estimation

Closest papers by meaning

• Inferring the Long-Term Causal Effects of Long-Term Treatments from Short-Term Experiments (ICML 2024 · similarity 0.86)
• CausalPFN: Amortized Causal Effect Estimation via In-Context Learning (NeurIPS 2025 · similarity 0.85)
• GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.85)
• A Practical Approach to Causal Inference over Time (AAAI 2025 · similarity 0.85)
• Marginal Causal Flows for Validation and Inference (NeurIPS 2024 · similarity 0.85)
• Do Contemporary Causal Inference Models Capture Real-World Heterogeneity? Findings from a Large-Scale Benchmark (ICLR 2025 · similarity 0.84)