Research map / Learning theory and algorithms / Prediction and decision losses

Treatment Effect Estimation for Optimal Decision-Making

NeurIPS 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Adaptive testing and experiments · 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.

Shows two-stage CATE meta-learners (e.g., DR-learner) can be suboptimal for thresholded decision-making, and proposes a retargeted policy-targeted CATE (PT-CATE) objective optimized with a three-step neural method with adaptive sigmoid indicator approximation.

Problem

Two-stage CATE estimators are optimal for CATE estimation but can yield suboptimal thresholded (treat if CATE>0) policies when the second-stage function class is restricted or regularized.

Approach

Retarget the second-stage CATE regression to a convex combination of CATE squared error and negative policy value of the thresholded policy, smoothing the indicator with an adaptive, covariate-dependent sigmoid scale alpha(X) learned as a region detector via a classification loss.

Key results

• On the Hillstrom dataset, PT-CATE with gamma=0.98 yields a 24.45% improved response probability over the gamma=0 CATE policy relative to the observational policy (policy value 1.805 vs 1.738).
• On synthetic settings, increasing gamma reduces policy loss compared with the gamma=0 two-stage baseline across PI, RA, IPW and DR pseudo-outcomes while only minimally increasing PEHE.
• Theorem shows that for any restricted neural class there exists a CATE for which the optimal CATE approximation yields a policy strictly worse than the best thresholded policy in the class.

Measured effects reported in the paper

Surprising findings

• The DR-learner-like two-stage CATE estimators that are optimal for CATE estimation can be strictly suboptimal for thresholded decision-making when the function class is restricted.

Limitations

• If the second-stage model class is unrestricted, the method yields no improvement over standard two-stage learners.
• Results on IPW and DR pseudo-outcomes are noisier due to propensity-score division variance.
• Choice of gamma relies on domain knowledge and sensitivity analysis; no principled automatic selection.

Experimental setup

Datasets: synthetic data (settings A and B) Hillstrom Email Marketing dataset

Benchmarks: synthetic settings A and B Hillstrom Email Marketing

Metrics: PEHE policy loss policy value improvement over observational policy

Baselines: PI-learner RA-learner IPW-learner DR-learner

Builds on: DR-learner X-learner R-learner Adam

Closest papers by meaning

• Unveiling the Potential of Robustness in Selecting Conditional Average Treatment Effect Estimators (NeurIPS 2024 · similarity 0.89)
• Learning Representations of Instruments for Partial Identification of Treatment Effects (ICML 2025 · similarity 0.88)
• AutoCATE: End-to-End, Automated Treatment Effect Estimation (ICML 2025 · similarity 0.88)
• Meta-Learners for Partially-Identified Treatment Effects Across Multiple Environments (ICML 2024 · similarity 0.87)
• A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.87)
• Bounds on Representation-Induced Confounding Bias for Treatment Effect Estimation (ICLR 2024 · similarity 0.86)