Research map / Probabilistic and scientific modeling / Treatment effects
Decision-Aware Proximal Bridge Learning for Optimal Treatment Selection
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 policy-targeted weighted bridge loss for proximal continuous-treatment selection, with a regret bound via weighted ill-posedness, instantiated as decision-aware variants of PMMR, NMMR, KPV, and DFPV.
Problem
Proximal bridge learning under hidden confounding is estimation-centric (global moment/surface accuracy), which can misallocate modeling effort away from the treatment region that determines the optimal continuous dose.
Approach
Reweight the bridge moment loss with a local-plus-global policy-targeted weight (kernel around the pseudo-optimal treatment plus inverse-density global term), iterated with response-surface projection and policy update; the weighted bridge loss is shown to control a regret surrogate via a weighted ill-posedness constant while preserving the population bridge target.
Key results
- On synthetic and semi-synthetic benchmarks, decision-aware weighting generally reduces policy regret for proximal bridge learners; DA-KPV has lowest regret on synthetic and DA-DFPV on semi-synthetic, with Wilcoxon tests showing differences for most models.
- Counterfactual RMSE shows no substantial difference between DA variants and baselines, so gains are concentrated in decision quality rather than global accuracy.
- Non-proximal S-learner baselines are consistently worse in regret, supporting the need for proxy-based adjustment under hidden confounding.
Surprising findings
- Decision-aware weighting often improves regret without corresponding improvements in global counterfactual RMSE or factual prediction error.
Limitations
- Relies on identification assumptions involving valid proxies, completeness, and bridge existence that are hard to verify in practice.
- Regret bound uses a weighted ill-posedness constant that is assumed finite; the global surface bound is intentionally coarse.
- Counterfactual RMSE gains are mixed, and DA-DFPV is sensitive to weighting hyperparameters.
Experimental setup
Datasets: fully synthetic proximal DGP (adapted from prior work) semi-synthetic TCGA gene-expression benchmark
Benchmarks: synthetic proximal DGP semi-synthetic TCGA benchmark
Metrics: policy regret counterfactual RMSE factual prediction error
Baselines: PMMR NMMR-V NMMR-U KPV DFPV S-learner weighted S-learner
Builds on: Outcome Oriented Counterfactual Prediction kernel proxy variables (KPV) proxy maximum moment restriction (PMMR) neural maximum moment restriction (NMMR) deep feature proxy variables (DFPV)
Closest papers by meaning
- Coupling Generative Modeling and an Autoencoder with the Causal Bridge (NeurIPS 2025 · similarity 0.86)
- An effective framework for estimating individualized treatment rules (NeurIPS 2024 · similarity 0.86)
- Doubly Robust Proximal Causal Learning for Continuous Treatments (ICLR 2024 · similarity 0.85)
- Optimistic Algorithms for Adaptive Estimation of the Average Treatment Effect (ICML 2025 · similarity 0.85)
- Treatment Effect Estimation for Optimal Decision-Making (NeurIPS 2025 · similarity 0.84)
- Off-policy estimation with adaptively collected data: the power of online learning (NeurIPS 2024 · similarity 0.84)
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.