Research map / Probabilistic and scientific modeling / Treatment effects
Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates
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 conditional counterfactual mean embeddings (CCME) in an RKHS with a doubly robust two-stage meta-estimator, instantiated as ridge, deep-feature and neural-kernel estimators, with finite-sample rates.
Problem
Estimating the full conditional distribution of counterfactual outcomes (heterogeneous treatment effects beyond scalar summaries) in a way that is robust to misspecified nuisance models and comes with finite-sample guarantees.
Approach
Debias the RKHS-valued outcome feature map with an AIPW-style pseudo-outcome (phi(Y) weighted by propensity and corrected by a first-stage CME), then regress the pseudo-outcome on the conditioning covariates with ridge, deep-feature, or neural-kernel regression; this yields double robustness in rates.
Key results
- Under correct specification DR matches IPW/PI for deep-feature and neural-kernel estimators and beats One-Step across all sample sizes (up to n=20000).
- When the outcome model is misspecified, PI variants fail to converge while DR remains consistent via the correct propensity score.
- On semi-synthetic MNIST, DR recovers canonical digit modes resembling the oracle, while One-Step shows bolder strokes from selection bias toward high-intensity images.
Surprising findings
- The Ridge Regression estimator's PI variant exhibits higher MSE than DR and IPW even under correct specification, attributed to selection bias.
- Estimates for covariate profile v2 underestimate the modes of the bimodal distribution even at n=20000.
Limitations
- Theory covers only the second stage in detail with first-stage error entering via min of nuisance risks; L4-type assumptions are replaced by boundedness.
- Experiments are on synthetic and a single semi-synthetic MNIST setup with small networks.
- Neural-Kernel estimator shows highest variance across seeds.
Experimental setup
Datasets: MNIST (semi-synthetic, noisy/treatment-assigned)
Benchmarks: synthetic bimodal counterfactual DGP semi-synthetic MNIST
Metrics: MSE against analytic ground-truth density empirical mode visualization
Baselines: One-Step estimator Inverse Propensity Weighting Plug-in estimator
Builds on: kernel mean embeddings doubly robust / AIPW estimation double machine learning conditional mean embedding counterfactual mean embeddings
Closest papers by meaning
- Doubly-Robust Estimation of Counterfactual Policy Mean Embeddings (NeurIPS 2025 · similarity 0.91)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.89)
- Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.88)
- A Dual-module Framework for Counterfactual Estimation over Time (ICML 2024 · similarity 0.88)
- Exogenous Matching: Learning Good Proposals for Tractable Counterfactual Estimation (NeurIPS 2024 · similarity 0.87)
- Learning Counterfactual Outcomes Under Rank Preservation (NeurIPS 2025 · similarity 0.87)
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.