Research map / Probabilistic and scientific modeling / Treatment effects
Learning Counterfactual Outcomes Under Rank Preservation
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 rank-preservation assumption for identifying counterfactual outcomes without a known SCM, plus a convex, kernel-based unbiased ideal loss for individual counterfactual estimation, evaluated on simulated, IHDP, and Jobs data.
Problem
Estimating individual-level counterfactual outcomes without a known structural causal model and without strong assumptions like homogeneity and strict monotonicity of the exogenous variable.
Approach
Replace homogeneity and strict monotonicity with a rank preservation assumption (Kendall rank correlation of 1 between factual and counterfactual outcomes given covariates), which identifies the counterfactual as the tau*-quantile of the counterfactual-arm distribution. This yields a convex ideal loss whose minimizer is the counterfactual, estimated with inverse propensity weighting and kernel smo
Key results
- On the simulated Sim-5 to Sim-40 datasets, the proposed method achieves the lowest in-sample and out-of-sample sqrt(PEHE) (e.g., 2.25 vs 2.80 for Quantile-Reg on Sim-10 in-sample).
- On IHDP, the proposed method reports sqrt(PEHE) of 1.41 in-sample and 1.50 out-of-sample, outperforming Quantile-Reg (1.43 and 1.56).
- The method is reported to be consistent and unbiased under known propensity scores, with bias depending only on propensity-score estimation error; it still outperforms baselines when rank preservation is violated in added experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed kernel-smoothed ideal-loss counterfactual estimator instead of Quantile-Reg | Individual treatment effect estimation · Sim-5 through Sim-40, in-sample and out-sample | √ε_PEHE: Sim-5 in-sample: 2.80 ± 0.06 -> 2.45 ± 0.17; Sim-5 out-sample: 2.54 ± 0.05 -> 2.28 ± 0.23; Sim-10 in-sample: 2.78 ± 0.08 | Helped |
| Use the proposed method rather than the strongest reported baseline | Individual treatment effect estimation · IHDP | √ε_PEHE: in-sample: 1.43 ± 0.05 -> 1.41 ± 0.02; out-sample: 1.56 ± 0.03 -> 1.50 ± 0.06 | Helped |
| Use the proposed method rather than the strongest reported baseline | Individual treatment effect estimation · IHDP in-sample | ε_ATE: 0.14 ± 0.09 -> 0.11 ± 0.10 | Helped |
| Use the proposed method rather than the strongest reported baseline | Individual treatment effect estimation · IHDP out-sample | ε_ATE: 0.15 ± 0.08 -> 0.13 ± 0.08 | Helped |
| Use the proposed method rather than the strongest reported baseline | Individual treatment effect estimation · Jobs out-sample | R_Pol: 0.15 ± 0.01 -> 0.11 ± 0.05 | Helped |
| Use the proposed method rather than Quantile-Reg | Individual treatment effect estimation · Jobs out-sample | ε_ATT: 0.07 ± 0.04 -> 0.05 ± 0.05 | Helped |
| Evaluate the proposed method with violated rank preservation | Individual treatment effect estimation · Sim-10 and Sim-40 with rank coefficient 0.3 or 0.5 | √ε_PEHE: null | Helped |
| Vary kernel function and bandwidth | Proposed method · Experiments with different kernels and bandwidths | √ε_PEHE | No clear effect |
Limitations
- Instability when propensity scores are extremely small in data-sparse settings
- Identification relies on the rank preservation assumption, which is a high-level condition not directly testable
- Estimator applies to discrete treatments; continuous extension is only sketched in the appendix
Experimental setup
Datasets: IHDP Jobs Simulated Sim-m dataset
Benchmarks: Simulated Sim-m datasets IHDP Jobs
Metrics: sqrt PEHE epsilon ATE epsilon ATT policy risk R_Pol
Baselines: T-learner X-learner BNN TARNet CFRNet CEVAE DragonNet DeRCFR DESCN ESCFR CFQP Quantile-Reg
Builds on: Quantile regression counterfactual estimation (Xie et al.) Pearl three-step counterfactual procedure Kendall rank correlation CFRNet
Closest papers by meaning
- Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models (ICLR 2024 · similarity 0.89)
- Doubly-Robust Estimation of Counterfactual Policy Mean Embeddings (NeurIPS 2025 · similarity 0.89)
- Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.88)
- Counterfactual Identification Under Monotonicity Constraints (AAAI 2025 · similarity 0.88)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.88)
- Causal Contrastive Learning for Counterfactual Regression Over Time (NeurIPS 2024 · similarity 0.88)
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.