Research map / Probabilistic and scientific modeling / Treatment effects
Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective
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 a counterfactual covering-radius bound for active-learning treatment effect estimation and two model-independent acquisition methods (greedy radius reduction and FCCM coverage maximization) that outperform AL and causal AL baselines on Toy, IBM and CMNIST.
Problem
Data-efficient treatment effect estimation under a limited labeling budget when treatment indicators and covariates are known in the pool but outcomes are costly to label; the PEHE risk is not directly optimizable because counterfactual outcomes are unobserved.
Approach
Derives a probabilistic PEHE upper bound in terms of factual and counterfactual covering radii, then selects samples via a greedy radius-reduction rule and a weighted-graph greedy coverage maximization (FCCM) that counts counterfactual coverage with weight alpha, allowing single-sample acquisition without paired queries.
Key results
- FCCM achieves a sqrt-PEHE gain of about 16-18% over the FCCM- ablation on Toy at 2/5 to 5/5 of budget (e.g., 0.7565 at full budget).
- On CMNIST, FCCM improves over FCCM- by roughly 24-30% across budget stages (1.6790 vs 2.2073 at full budget).
- On IBM the ablation gain is near zero (-1% to 0%) because treated and control distributions overlap heavily.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add counterfactual coverage maximization to FCCM- (ablation) | FCCM · Toy | sqrt PEHE: +0%, +17%, +18%, +18%, +16% | Helped |
| Add counterfactual coverage maximization to FCCM- (ablation) | FCCM · IBM | sqrt PEHE: -1%, -1%, -1%, +0%, +0% | Mixed |
| Add counterfactual coverage maximization to FCCM- (ablation) | FCCM · CMNIST | sqrt PEHE: +24%, +30%, +28%, +25%, +24% | Helped |
| Use FCCM's joint factual and counterfactual coverage acquisition instead of the compared baselines | Treatment effect estimation · Toy, IBM, and CMNIST | sqrt PEHE | Helped |
Surprising findings
- BAIT, a general AL method, outperforms many treatment-effect-specific AL baselines on CMNIST but underperforms on Toy and IBM.
- Current DTEE baselines do not consistently outperform general AL methods across datasets.
Limitations
- FCCM degrades when treatment groups have non-overlapping regions in raw feature space, since no counterfactual pairs exist.
- Theoretical bound relies on Lipschitz continuity and a constant kappa assumption; zero training loss assumption.
- Experiments use simulated and semi-synthetic data only.
Experimental setup
Datasets: Toy (simulated 2D) IBM (Linked Births and Infant Deaths Database semi-synthetic) CMNIST
Benchmarks: Toy IBM CMNIST
Metrics: sqrt PEHE performance gain relative to FCCM-
Baselines: BADGE BAIT LCMD QHTE Causal-BALD MACAL muBALD rhoBALD murhoBALD
Builds on: ProbCover core-set k-Center QHTE MACAL Causal-BALD DUE
Closest papers by meaning
- Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing (AAAI 2024 · similarity 0.87)
- Adversarially Balanced Representation for Continuous Treatment Effect Estimation (AAAI 2024 · similarity 0.86)
- A Generative Approach for Treatment Effect Estimation under Collider Bias: From an Out-of-Distribution Perspective (ICML 2024 · similarity 0.85)
- KG-TREAT: Pre-training for Treatment Effect Estimation by Synergizing Patient Data with Knowledge Graphs (AAAI 2024 · similarity 0.85)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.85)
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.