🤖 machine learning

Decision-Aware Proximal Bridge Learning for Optimal Treatment Selection

This paper introduces a decision-aware proximal bridge learning framework that employs a policy-targeted weighted loss to optimize individualized treatment selection under hidden confounding, thereby reducing regret by prioritizing modeling effort in decision-relevant regions rather than uniformly across the entire action space.

Original authors: Tomàs Garriga, Alejandro Almodóvar, Axel Brando, Gerard Sanz, Eduard Serrahima de Cambra, Juan Parras

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tomàs Garriga, Alejandro Almodóvar, Axel Brando, Gerard Sanz, Eduard Serrahima de Cambra, Juan Parras

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Finding the "Sweet Spot" in a Foggy Room

Imagine you are a doctor trying to prescribe the perfect dose of a new medicine to a patient. You have a lot of data about past patients, but there is a catch: you can't see everything. There are hidden factors (like a patient's secret stress levels or unrecorded genetic quirks) that influence both how much medicine they were given and how they got better. In statistics, we call this "hidden confounding."

Because of this fog, you can't just look at the data and say, "Oh, people who took 5mg got better, so 5mg is the answer." That might be wrong because the people who took 5mg might have been healthier to begin with.

To solve this, scientists use a trick called Proximal Causal Inference. Think of it like using two "spies" to figure out what's happening in the foggy room:

Spy Z (The Treatment Spy): This spy watches what factors influenced the doctor's decision to give a specific dose.
Spy W (The Outcome Spy): This spy watches what factors influenced how the patient recovered.

By comparing these two spies, we can mathematically "bridge" the gap and figure out the true cause-and-effect relationship, even without seeing the hidden factors directly.

The Old Way: Mapping the Whole Territory

Traditionally, when scientists use these "spies" to build a model, they try to draw a perfect map of the entire territory. They want to know exactly what happens if you take 1mg, 2mg, 3mg, all the way to 100mg. They try to make the map accurate everywhere, equally.

The Flaw: This is like trying to draw a perfect map of a whole country just to find the best spot for a single picnic. You might spend all your effort making the map of the desert (where no one goes) incredibly detailed, while the map of the picnic spot (the optimal dose) is a little blurry. If the map is blurry right where you need to stand, you might pick the wrong spot for your picnic.

The paper argues that for decision-making, global accuracy is a waste of energy. We don't care if the model is perfect for doses that are clearly too high or too low. We only care that the model is super accurate right around the "sweet spot" (the optimal dose) for each individual.

The New Solution: The "Decision-Aware" Bridge

The authors propose a new way to build this bridge. Instead of trying to be perfect everywhere, they make the model decision-aware.

The Analogy: The Flashlight in the Dark
Imagine you are in a dark room trying to find a specific key on the floor.

The Old Method: You turn on a dim light that illuminates the whole room evenly. You might miss the key because the light is too spread out.
The New Method: You use a flashlight. You start by guessing where the key might be. You shine the bright beam of the flashlight right on that spot to get a super-clear view. If you realize you were wrong, you move the flashlight to the new spot and focus there.

In technical terms, the paper introduces a "weighted bridge loss."

The Pilot: The model makes a first guess at the best dose (the "pseudo-optimal" dose).
The Weighting: The model then says, "I don't care as much about being accurate for doses far away from this guess. I will focus all my energy on getting the math right right next to this guess."
The Loop: It updates the dose, moves the "flashlight," and refines the guess again.

Why This Matters (The "Bridge" Part)

The paper proves mathematically that this "flashlight" approach works even in the "foggy room" (with hidden confounders).

They show that by focusing the math on the decision-relevant areas, you can actually reduce regret. In this context, "regret" is the difference between the outcome you got and the outcome you could have gotten if you had picked the perfect dose.

Old Way: Low regret? Maybe. But only if the whole map was perfect.
New Way: Low regret? Yes, because the map is perfect exactly where it counts.

The Results: Does the Flashlight Work?

The authors tested this idea on two types of "simulated worlds":

A totally made-up world (Synthetic): Where they knew the ground truth.
A semi-real world (TCGA): Based on real cancer gene data, but with the treatment and outcomes simulated.

What they found:

The "Decision-Aware" models (the ones with the flashlight) consistently made better treatment choices than the standard models. They reduced the "regret" (the cost of picking a suboptimal dose).
Interestingly, the Decision-Aware models didn't necessarily get better at predicting outcomes for every dose. Sometimes, they were actually slightly worse at predicting the "useless" doses (like 99mg). But that didn't matter! Because they were better at the right dose, the final decision was superior.

Summary

This paper is about changing how we teach computers to make medical (or policy) decisions when data is messy and hidden factors exist.

Old approach: "Let's build a perfect map of the whole world so we can find the best spot."
New approach: "Let's build a map that is blurry everywhere except right where we think the best spot is, and keep moving that focus until we find the true best spot."

By focusing their energy on the "decision-relevant" regions rather than the whole space, the authors show that we can make better individualized treatment choices, even when we can't see all the hidden variables.

Technical Summary: Decision-Aware Proximal Bridge Learning for Optimal Treatment Selection

Problem Statement

The paper addresses the challenge of individualized treatment selection for continuous actions (e.g., drug dosage) in the presence of unmeasured confounding. While standard causal inference methods often focus on estimating the global causal response surface (dose-response curve) with uniform accuracy, the authors argue that this objective is misaligned with the ultimate goal of decision-making. A model may fit the response surface well on average but remain inaccurate in the specific regions near the optimal treatment dose, leading to suboptimal policy selection.

This problem is particularly acute in proximal causal inference, a framework designed to handle hidden confounding using proxy variables (treatment-inducing $Z$ and outcome-inducing $W$ ) and bridge functions. Existing proximal methods (e.g., PMMR, NMMR, KPV, DFPV) are primarily estimation-centric, optimizing for global accuracy of the bridge function or the causal surface. They do not explicitly account for the fact that errors in suboptimal treatment regions have little impact on policy value, whereas small errors near the optimum can drastically alter the recommended treatment.

Methodology

The authors propose a Decision-Aware Proximal Bridge Learning framework that shifts the optimization objective from global surface accuracy to policy-targeted regret minimization.

1. Theoretical Foundation: Weighted Bridge Loss

The core theoretical contribution is the formulation of a policy-targeted weighted bridge loss. The authors recognize that proximal bridge estimation is an inverse problem. They introduce a weighted norm for solving this problem that emphasizes treatment regions relevant to the induced policy while retaining a global component for stability.

Regret Decomposition: The paper establishes that controlling policy regret requires controlling the surface error near two specific points: the oracle optimal dose ( $a^*(X)$ ) and the learned pseudo-optimal dose ( $\pi_h(X)$ ).
Surface Surrogate: A combined surface loss is constructed, consisting of:
- A local component ( $A_\tau$ ): A kernel-weighted loss around the pseudo-optimal treatment $\pi_h(X)$ , which approximates the error at the decision boundary.
- A global component ( $B$ ): A uniform loss over the entire treatment space to stabilize the inverse problem and protect against missing the oracle region.
Weighted Ill-Posedness: The authors prove that minimizing a weighted bridge moment loss controls the weighted surface loss (and thus a regret surrogate) under a weighted ill-posedness condition. This condition ensures that the inverse problem remains stable even when the loss is reweighted.
Target Preservation: Crucially, the paper proves that as long as the weights are strictly positive, the population minimizer of the weighted bridge loss remains the true bridge function $h_0$ . The weighting changes the geometry of the optimization (penalizing deviations more heavily in decision-relevant regions) but does not alter the target population parameter.

2. Algorithmic Implementation

The framework is instantiated as an iterative algorithm that alternates between four steps:

Bridge Estimation: Fit a bridge function $h$ using a weighted empirical bridge-moment loss. The weights are derived from the current policy.
Response Surface Projection: Project the fitted bridge into an induced causal response surface $m_h(a, x) = E[h(a, W, x) | X=x]$ . This is done via cross-fitting on a treatment grid to generate pseudo-outcomes.
Policy Update: Derive a new pseudo-optimal policy $\pi_h(x)$ by maximizing the estimated response surface.
Weight Refinement: Update the weighting function $\omega(a, x)$ for the next iteration. The new weights combine a global inverse-propensity term with a localized term centered at the current pseudo-optimal treatment $\pi_h(x)$ .

This approach acts as a general wrapper applicable to various proximal solvers. The authors demonstrate its implementation for:

PMMR (Proxy Maximum Moment Restriction)
NMMR (Neural Maximum Moment Restriction)
KPV (Kernel Proxy Variables)
DFPV (Deep Feature Proxy Variables)

In each case, the weights are integrated into the solver's objective function (e.g., by scaling residuals in MMR-based methods or as sample weights in second-stage regressions for two-stage methods).

Key Contributions

Problem Formulation: The paper formulates individualized continuous-treatment selection under hidden confounding as a regret-targeted proximal bridge-learning problem, extending the concept of outcome-oriented counterfactual prediction to the proximal setting.
Theoretical Guarantees: It provides a regret bound showing that the proposed weighted bridge loss controls a regret-oriented surface surrogate through a weighted ill-posedness constant. This demonstrates that decision-aware weighting is theoretically sound for controlling treatment-selection regret.
Algorithmic Framework: The authors propose an iterative algorithm that alternates between bridge estimation, surface projection, policy update, and weight refinement, instantiating it for four distinct proximal bridge solvers.
Empirical Validation: The method is evaluated on synthetic and semi-synthetic benchmarks (including a TCGA gene-expression dataset) with hidden confounding and proxy variables.

Results

Empirical evaluation on synthetic and semi-synthetic datasets yields the following findings:

Regret Reduction: Decision-aware (DA) weighting consistently reduces policy regret across several proximal bridge solvers (PMMR, NMMR, KPV, DFPV) compared to their standard, unweighted counterparts. The gains are particularly pronounced for solvers whose baseline response surfaces exhibit significant errors near the treatment optimum.
Counterfactual Estimation: The paper reports mixed results regarding global counterfactual estimation error (RMSE). The decision-aware variants do not uniformly improve global counterfactual accuracy, which is expected given that the objective is not designed to minimize error uniformly across the treatment space. In some cases, DA variants show no significant difference in global RMSE compared to baselines.
Robustness: The method demonstrates robustness to variations in hyperparameters (localization bandwidth $\tau$ , strength $\lambda$ , and number of rounds). It also shows that non-proximal baselines (e.g., S-Learner) fail significantly as confounding strength increases, highlighting the necessity of proxy-based adjustment.
Individual-Level Improvement: Case studies illustrate that DA methods improve the response estimate specifically near the oracle optimum, leading to treatment recommendations closer to the optimal dose, even if the full response surface is not uniformly improved.

Significance and Claims

The paper claims that aligning bridge estimation with the downstream treatment-selection objective can improve decision quality in proximal causal inference, even when global response estimation is not uniformly improved.

Distinction from Prior Work: Unlike outcome-oriented methods in unconfounded settings, this work addresses the specific challenge of unmeasured confounding, where the causal surface is only accessible via a proximal bridge equation. The decision-targeting is imposed at the level of the inverse conditional-moment problem rather than a supervised prediction loss.
Practical Utility: The framework offers a practical, general-purpose wrapper that can be applied to existing proximal solvers to enhance their performance in decision-making tasks without requiring changes to the underlying identification assumptions.
Limitations: The authors modestly acknowledge that the approach relies on identification assumptions involving valid proxies and bridge existence, which may be difficult to verify in practice. The method does not claim to solve the fundamental identification problem but rather optimizes the estimation process for a specific decision-theoretic goal once identification is assumed possible.

In summary, the paper argues that in proximal causal inference, global accuracy is not the right target for treatment selection. By reweighting the bridge loss to focus on decision-relevant regions while maintaining global stability, the proposed method achieves lower policy regret, offering a more effective approach for individualized treatment selection under hidden confounding.