SHIFT: Robust Double Machine Learning for Average Dose-Response Functions under Heavy-Tailed Contamination

2026-05-04 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

SHIFT (Self-calibrated Heavy-tail Inlier-Fit with Tempering) is a robust Double Machine Learning (DML) estimator designed to address "functional smearing" in Average Dose-Response Function (ADRF) estimation, particularly under heavy-tailed contamination. Traditional DML pipelines using kernel-weighted local-linear smoothers suffer from unbounded functional influence, where a single outlier can bias the entire dose-response curve. SHIFT combines cross-fit nuisance orthogonalization with a kernel-local Welsch-loss second stage, optimized via Graduated Non-Convexity (GNC). Its key innovation is a defensive Ordinary Least Squares (OLS) refit that uses the Median Absolute Deviation (MAD) of post-GNC residuals for its inlier cutoff, rather than the raw-outcome MAD. This architectural choice reduces level-RMSE from 1.03 to 0.33 on localized-contamination stress tests at $p=0.25$, while maintaining performance on clean data. SHIFT also provides a non-uniform per-sample weight vector, achieving mean F1-scores of approximately 0.96 for outlier detection on Gaussian-jump DGPs. The estimator is paired with a six-technique Extreme Value Theory (EVT) diagnostic suite to help practitioners characterize tail behavior and choose appropriate robust methods.

Key takeaway

For AI Engineers and Research Scientists working on causal inference with continuous treatments, SHIFT offers a robust solution for Average Dose-Response Function estimation, especially when dealing with contaminated observational data. You should integrate SHIFT into your DML pipelines to improve shape recovery and gain a native outlier mask, particularly in scenarios with localized contamination. Additionally, leverage the accompanying EVT diagnostic suite to empirically understand the tail characteristics of your data, guiding your choice between SHIFT, Huber-DML, or Quantile-DML for optimal performance.

Key insights

SHIFT robustifies DML for ADRF by using post-GNC residual MAD for outlier detection, improving accuracy and providing an outlier mask.

Principles

• Redescending M-estimators mitigate unbounded functional influence.
• Post-GNC residual MAD improves outlier detection in localized contamination.
• EVT diagnostics inform robust estimator selection based on tail characteristics.

Method

SHIFT employs cross-fit nuisance orthogonalization, kernel-local GNC with Welsch loss, and a defensive OLS refit using post-GNC residual MAD for inlier cutoffs.

In practice

• Use SHIFT for robust ADRF estimation with outlier detection.
• Employ EVT diagnostics to classify tail behavior (e.g., Fréchet vs. Weibull).
• Consider linear nuisance models for robust DML under uniform contamination.

Topics

• Robust Double Machine Learning
• Average Dose-Response Function
• Heavy-Tailed Contamination
• Graduated Non-Convexity
• Extreme Value Theory Diagnostics

Code references

• EichiUehara/ADRF-Robust-DML
• uber/causalml
• py-why/EconML

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Bulk-Calibrated Credal Ambiguity Sets: Fast, Tractable Decision Making under Out-of-Sample Contamination — Bulk-calibrated credal ambiguity sets enable tractable, finite worst-case risk in Huber contamination models for robust decision-making.
• A Decision-Theoretic View of Test-Time Training: When, How Far, and Which Directions to Adapt — TTT's effectiveness hinges on spectrally matching update steps and aligning update directions with query relevance.
• Tail-Shape Estimation in LLM Evaluation Is Fragile: A Protocol for Diagnosing False Positives — Tail-shape estimation in LLM evaluation is fragile, necessitating a rigorous diagnostic protocol to prevent false positives.
• A Deployment Audit of Release-Side Risk in Conformal Triage under Prevalence Shift — A leakage-aware audit improves conformal triage safety by identifying event-positive releases under prevalence shift.
• Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation — D³-Net combines SDR-based training with LTMLE targeting to robustly estimate longitudinal treatment effects, mitigating error propagation.
• The Reliability Gap in Benchmark Auditing: Distribution Shift and Scale as Failure Modes of Contamination Detection — Current statistical methods for detecting LLM benchmark contamination are unreliable in realistic auditing due to distribution shift and scale.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.