acceptodds
Under review as a conference paper at ICLR 2027

Proximal Policy Learning with Nonunique Outcome Confounding Bridges

Abstract

Learning treatment policies from observational data with unmeasured confounding requires valid comparisons of candidate policies. Proximal causal inference uses proxy variables and outcome confounding bridges for this purpose, but the observed bridge equation can admit multiple solutions. Existing results on nonunique bridges show that some causal functionals remain identified, yet they do not resolve policy choice when admissible bridges imply different rankings. In this paper, we characterize when policy value contrasts are invariant over the observed bridge solution set and, when they are not, formulate minimax regret over the joint bridge solution set. We derive an observed-moment dual and a convex empirical program for randomized mixtures of a fixed finite collection of policies. We establish a finite-sample excess-regret bound and give conditions under which a finite dual radius exactly represents the population objective. Under finite-feature rank and interior conditions, this exact representation yields root- excess worst-case regret at zero bridge residual tolerance for a fixed finite policy hull when the global optimization error is root-. We also relate worst-case regret over observed bridge solutions to worst-case regret over compatible causal models. Controlled experiments show the decision effects of selecting one bridge and relaxing joint restrictions, and quantify the tradeoff between minimizing worst-case regret and maximizing worst-case policy value.

Primary area:
causal reasoning
Keywords:
Treatment Effect, Causal Inference, Proximal Policy Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6cn4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies learning a policy with unobserved confounding. In the proxy/bridge learning setting, it is assumed that there is some observed "proxy" variable for the hidden confounding, W, which affects only the treatment and not the outcome. The "bridge equation" looks for a (unique) function of treatment A, W, and covariates X (only) that exactly matches the observed outcomes (which depend on the full hidden confounding, U). This paper broadens this framework to consider a set of bridges that are not unique and/or may only approximately match the observed outcomes. In that case, we need to be able to compare policies whose predicted outcome may depend on the bridge model used. The paper gives a concrete procedure for comparing policies by minimizing regret across feasible bridge models, and a tractable convex optimization to search for policies that minimize worst case regret.

Critical strengths

Learning policies with unobserved confounding is an important problem. The hard part is generally to come up with setting / assumptions that are realistic enough to be useful, but still allow for non-trivial results. I thought the setup proposed sounded intuitive and useful. Previous work considered that a proxy for hidden confounding that affects only the treatment may be available - but only if there is an identifiable "bridge" model that connects the proxy and outcomes. What I like is that this work doesn't assume the (unrealistic) case that we get perfect identifiability of the bridge equation. That seems much more useful, allowing us to use imperfect confounder proxies to still decrease our regret in policy learning.

I'm familiar with causal inference in general, but not this particular literature, but I found the writing and setup to be clear and interesting. The notation used to define everything looked logical.

Experiments using semi-synthetic data validate the approach and test the effects of different bridges and the impact of the convexification.

Critical weaknesses

I did not notice any critical weaknesses, but I'm not familiar with this literature so did not follow everything in Sec. 5.

Minor comments

No need to discuss these or even take them into account, as some of this may just be due to conventions of the sub-field.

  • I would have liked a little more concrete motivation in the intro. E.g. "We want to test policy for drug A, but people who have access in the study may have more resources and this affects outcome - hence W could represent socioeconomic proxy (zip code, income)." I realize, though, that it is surprisingly not common in this type of literature to do this. And, maybe the previous proxy works have covered this so it seems unnecessary.
    1. I didn't know what the cond. moment equation was. Later it is stated that means Eq. 1 - I guess that wasn't obvious to me from the form.
    1. that Pi - Pi notation looks strange.
Questions for authors

No major questions, see above. I'll read the other reviews and rebuttals and take these into account in final review.

8 comments
Reviewer_6cn4OpenReview ↗

Thoughts on rev. 4qwy discussion

I'm following the discussion with 4qwy, who declared high confidence for this review.

  • The decision target depends on an untestable, user-chosen bridge class.

I didn't find this argument the most compelling - all of causal inference is about finding minimal, reasonable assumptions that allow for causal conclusions. Intuitively, the assumptions seemed reasonable to me, and the authors do propose a way to test the impact of this.

  • The finite-sample bounds are uninformative, and exact bridge sets are often empty

This seemed more interesting - if the bounds are vacuous then the results may be useless. From the rebuttal, it seems that it is true that at practical sample sizes the bounds are loose, but with enough samples they could become tight. This makes me a little ambivalent, though I can imagine that rankings are meaningful even if bounds are loose, though I don't recall anything like that being tested. And it does change the impact of the contribution.

  • The experiments are small, purpose-built simulations with no established baselines.

If the bounds are quite loose, some empirical causal benchmarks might be useful, I agree.

For [Q1], I haven't read this literature, so hope to see if 4qwy agrees with the author response clarifying the contribution.

AuthorsOpenReview ↗

Response to Reviewer 6cn4 (1/3)

Thank you for recognizing the value of policy learning with nonunique bridges and the clarity of the formulation! We address the concrete motivation, conditional moment equation, and policy notation below.

[M1] A concrete motivation from RHC

The added Right Heart Catheterization (RHC) application illustrates why different admissible bridges can lead to different treatment recommendations.

1. The application assigns RHC probabilities by age group using observed proxies for unmeasured illness severity.

The cohort contains 5,735 critically ill patients (Connors et al., 1996). Treatment AA indicates first-day RHC, outcome YY is survival truncated at 30 days and divided by 30, and XX contains deployment covariates. Unmeasured illness severity UU may affect both RHC use and survival, confounding a direct treatment comparison.

Following Cui et al. (2024), we use blood pH and hematocrit as outcome-inducing proxies WW, and oxygenation and carbon dioxide measurements as treatment-inducing proxies ZZ. Policies choose separate treatment probabilities for patients younger than 65 and those aged 65 or older.

2. If admissible bridges rank these policies differently, selecting one bridge can change the recommendation.

Following your suggestion, the revised Introduction uses this example to explain the decision problem concretely.

Thus, our method addresses the uncertainty in treatment recommendations by retaining the joint bridge restrictions and minimizing worst-case regret.

AuthorsOpenReview ↗

Response to Reviewer 6cn4 (2/3)

[M2] The conditional bridge equation and proxy roles

The bridge equation equates conditional means, so multiple bridges can satisfy it without reproducing each individual's outcome.

Please allow us to clarify the equation and the proxy roles.

1. The equality is conditional on observed treatment, treatment-inducing proxies, and deployment covariates.

In Section 3.1, Th=E[hA(W,X)∣A,Z,X]Th=\mathbb E[h_A(W,X)\mid A,Z,X] and m=E[Y∣A,Z,X]m=\mathbb E[Y\mid A,Z,X], so Th=mTh=m means

E[Y−hA(W,X)∣A,Z,X]=0,\mathbb E[Y-h_A(W,X)\mid A,Z,X]=0,

Here, hAh_A is the bridge for observed treatment AA. The equation requires equal means conditional on A,Z,XA,Z,X, not Y=hA(W,X)Y=h_A(W,X) for every individual. Several bridges can satisfy the same equation exactly.

2. WW is the outcome-inducing proxy and is not assumed to affect only treatment.

Assumption 1 requires

Y⊥Z∣(A,U,X),W⊥(A,Z)∣(U,X),Y\perp Z\mid(A,U,X), \qquad W\perp(A,Z)\mid(U,X),

where ⊥\perp denotes conditional independence. Section 3.1 now spells out the conditional-mean equation alongside the stated proxy roles and assumptions.

Thus, several bridges can satisfy the same conditional mean equation under the stated proxy assumptions.

Reviewer_uAUSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies policy learning when the proximal outcome bridge equation admits multiple solutions and policy comparisons may be unidentified. It proposes a minimax-regret framework over the joint bridge solution set, derives an observed-moment dual and a convex empirical learner, and establishes finite-sample regret guarantees. Synthetic and semi-synthetic experiments examine bridge selection, joint restrictions, and decision criteria.

Critical strengths
  1. The paper addresses policy learning under nonunique proximal bridges and characterizes when policy contrasts remain identifiable, extending the discussion beyond identification of individual causal functionals.

  2. The observed-moment dual and convex formulation provide a tractable minimax-regret learner, with finite-radius exact representation and root-n excess-regret guarantees under the stated conditions.

  3. The experiments isolate the effects of bridge selection, joint restrictions, and decision criteria, while reporting both bridge-based regret and generating-world causal loss.

Critical weaknesses
  1. The sufficient dual radius depends on unknown population quantities, while the experiments rely on information from the generating laws, leaving practical parameter selection unresolved.

  2. The framework assumes that the true causal bridge belongs to a prespecified function class, without establishing guarantees when this assumption is violated.

  3. The theoretical guarantee relies on finite-feature rank and interior conditions, and its practical usefulness under weak proxies or ill-conditioned moment operators remains unclear.

  4. The experiments rely primarily on synthetic and semi-synthetic settings with simple policy classes, providing limited evidence of effectiveness in realistic applications.

Questions for authors
  1. Can the empirical learner and its theoretical guarantees be extended beyond finite-dimensional linear representations to infinite-dimensional function spaces such as RKHS?

  2. Under what practical conditions do the assumptions required for the root-n regret guarantee hold, particularly with weak proxies or ill-conditioned moment operators?

  3. Can the dual radius and bridge radius be selected using only observed data, and how sensitive is the learned policy to these choices?

7 comments
AuthorsOpenReview ↗

Response to Reviewer uAUS (1/4)

Thank you for recognizing the contribution of the observed-moment dual, convex learner, and regret guarantees! We address the RKHS extension, misspecification of the bridge class, weak proxies, and practical parameter selection with additional experiments below.

[Q1] Extension to RKHS bridges

For RKHS bridges, the empirical problem reduces to a finite convex program, and we can bound its sampling error. The population regret bound also includes representation error.

The reproducing kernel Hilbert space (RKHS) formulations in Appendices C.2 and C.5 provide these results. Exact population representation additionally requires suitable operator conditions.

1. Sample kernel representations reduce empirical optimization to a finite-dimensional convex program.

Place bridges and dual test functions in RKHSs with kernels k,ℓk,\ell. For observations Si=(Ai,Zi,Xi)\mathsf S_i=(A_i,Z_i,X_i), comprising treatment, treatment-inducing proxies, and deployment covariates, projecting a dual function onto the sample kernel sections preserves its sample values and cannot increase its norm. Thus,

f(⋅)=∑i=1nαiℓ(Si,⋅),f(\cdot)=\sum_{i=1}^n\alpha_i\ell(\mathsf S_i,\cdot),

where αi\alpha_i are optimization variables. Empirical bridge moments have analogous representations. Consequently, a fixed finite policy hull gives a finite-dimensional convex second-order cone program in the sample kernel matrices. Positive semidefinite square roots compute the norms even for singular matrices, without matrix inversion.

2. Bounded kernels and outcomes give root-n sampling control for fixed function classes and norm bounds, with representation and optimization errors retained.

Theorem 7 in Appendix C.5 gives

Rρ(π^)−Rρ∗≤aρ+2en+eopt.R_\rho(\widehat\pi)-R_\rho^* \le a_\rho+2e_n+e_{\mathrm{opt}}.

Here, RρR_\rho is worst-case regret over the bridge set with residual tolerance ρ\rho, Rρ∗R_\rho^* is its minimum over the policy class, and π^\widehat\pi is the learned policy. The terms are population representation error aρa_\rho from restricting the dual space and radius, sampling error ene_n, and global optimization error eopte_{\mathrm{opt}}.

With bounded kernel diagonals and outcomes, fixed function classes and norm bounds, and a fixed finite policy hull, en=OP(n−1/2)e_n=O_P(n^{-1/2}) also holds in infinite-dimensional RKHSs. This uses a fixed computational norm, while Theorem 2 additionally handles an estimated L2L^2 norm. Therefore, if aρ=0a_\rho=0 and eopt=OP(n−1/2)e_{\mathrm{opt}}=O_P(n^{-1/2}), the excess-regret rate is root-nn.

Section 4.2 now summarizes finite kernel computation and the retained error terms, with details in Appendices C.2 and C.5.

Thus, the computation and sampling theory extend to RKHSs, while the representation term specifies what is needed for a population excess-regret rate.

AuthorsOpenReview ↗

Response to Reviewer uAUS (2/4)

[W2] Bridge-class misspecification

The learning guarantee concerns the declared bridge objective, and its causal interpretation additionally requires a causal bridge within that class.

Your question helps make the scope of these two statements explicit.

1. Under the feasibility and representation conditions, the learning guarantee still applies to the specified bridge objective.

The learner optimizes over the specified bridges, and the statistical guarantee concerns the best policy for that objective. If the class excludes the true causal bridge, the causal-regret bound cannot be invoked.

2. Extending the causal guarantee requires controlling how bridge-class misspecification changes policy contrasts.

Observed moments may leave null-space components undetermined, so fit, empirical feasibility, and validation error cannot establish causal-bridge inclusion. Extending the guarantee would require specifying the form and magnitude of misspecification and bounding its effect on policy comparisons.

Section 3.1 now makes causal-bridge inclusion explicit, and the Limitations paragraph identifies structured bridge-class misspecification as future work.

Thus, the theorem retains a decision guarantee for the declared model, while causal guarantees under misspecification require control of the resulting contrast errors.

AuthorsOpenReview ↗

Response to Reviewer uAUS (3/4)

[W3, Q2] Learning with weak proxies and rank-deficient operators

The root-nn guarantee allows nonunique bridges and rank-deficient operators, but its constants depend on moment strength and an interior bridge solution.

Theorem 3 and Corollary 8 in Appendix C.8 specify the conditions.

1. A complete moment representation, positive rank, and a strict interior bridge solution give a sufficient finite dual radius for exact representation.

Write the population moment equations as MT⊤θ=vM_T^\top\theta=v, where θ\theta contains bridge coefficients, MTM_T contains cross-moments of bridge and test features, and vv contains the expected products of the outcome and the test features. The bridge radius imposes ∣θ∣2≤B\\|\theta\\|_2\le B, and Λ\Lambda bounds dual coefficients.

  • The finite moment equations must represent the full conditional bridge equation within the bridge class. This holds if conditional residuals lie in the span of the test features, as verified for the experimental designs in Appendix D.4.4.
  • MTM_T must have positive rank and smallest nonzero singular value s\*>0s_\* > 0, while its null space may still change policy rankings. Thus, full rank and unique policy rankings are not required.
  • The minimum-norm bridge solution must lie strictly inside the radius-BB ball. The condition ∣θmin⁡∣2<B\\|\theta_{\min}\\|_2 < B gives margin a=B2−∣θmin⁡∣22>0a=\sqrt{B^2-\\|\theta_{\min}\\|_2^2} > 0.

Let KΠK_\Pi be the largest norm of a vector representing a policy contrast in the space of bridge coefficients over class Π\Pi. A sufficient radius is

Λ0=BKΠs\*a.\Lambda_0=\frac{BK_\Pi}{s_\*a}.

Under these conditions and the boundedness and measurability assumptions, fix the distribution, features, finite policy hull, and norm bounds, with Λ≥Λ0\Lambda\ge\Lambda_0 and ρ=0\rho=0. For independent, identically distributed observations and eopt=OP(n−1/2)e_{\mathrm{opt}}=O_P(n^{-1/2}),

R0(π^)−R0∗=OP(n−1/2).R_0(\widehat\pi)-R_0^*=O_P(n^{-1/2}).

2. Smaller nonzero singular values or interior margins increase the sufficient dual radius and sampling bound, so root-nn is a fixed-distribution guarantee.

Empirical singular values, interior margins, and stability of treatment probabilities can reveal unstable settings, but do not verify population conditions. The bound constants identify why weak proxies can require more data. Sharper numerical guarantees as the operator approaches singularity require further analysis.

3. Weaker proxies increase the sufficient radius in the existing experiment, while the learner maintains low regret for the same population target.

Appendix E.1 varies ss, the strength of the treatment-inducing proxy's association with one latent confounder, over 0.25,0.5,0.750.25,0.5,0.75. Each setting uses 200 datasets of size 4,000, policies with constant treatment probabilities, B=2B=2, ρ=0\rho=0, and Λ=125\Lambda=125. The exact population bridge set is the same at all three strengths. The sufficient values of Λ\Lambda are 42.574,21.407,14.42042.574,21.407,14.420, respectively (Appendix D.2). Thus, weakening the proxy raises the sufficient radius while leaving the decision target unchanged.

At s=0.25s=0.25 and 0.750.75, mean excess bridge regret is 0.008160.00816 and 0.006770.00677, respectively, compared with 0.053350.05335 for minimum-norm bridge selection. All 600 empirical bridge sets are nonempty and all fits complete. Together, these results connect the operator-dependent sufficient radius with learning performance over the tested strength range.

Section 5.2 explains how the sufficient radius depends on the operator. Section 6.2 and Appendix E.1 connect this dependence to the existing proxy-strength experiment.

Thus, weak proxies affect the constants, while the tested strengths still permit learning of a common bridge target under the stated conditions.

Reviewer_4QWYOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how to learn treatment policies from observational data when unmeasured confounding is present and the proximal outcome bridge equation has multiple valid solutions. Since different bridge solutions may lead to different policy choices, the authors first identify when policy comparisons remain unchanged across these solutions. When the comparisons are uncertain, they propose a minimax-regret approach that selects a policy by minimizing its worst-case loss relative to the best policy under the same bridge. The authors develop a convex optimization method for mixtures of a fixed set of policies and provide theoretical guarantees, including a finite-sample regret bound and a root-nn convergence rate under certain assumptions. Experiments on synthetic and semi-synthetic ACIC data compare the proposed method with minimum-norm bridge selection, a method that ignores joint restrictions, and maximin policy value optimization.

Critical strengths
  • The paper addresses an important problem in proximal causal policy learning: how to select policies when multiple admissible bridges imply different treatment decisions. It clearly distinguishes between invariance of policy contrasts, preservation of their signs, and the existence of a common optimal policy.
  • The proposed method evaluates candidate and comparator policies under the same bridge, preserving dependencies across policy comparisons rather than treating marginal contrast intervals independently. Figure 1(c,d) clearly illustrates how ignoring these joint restrictions can change the learned policy.
  • The paper provides a solid theoretical analysis. It develops a convex optimization method for learning policies and establishes finite-sample regret bounds. Theorem 3 gives sufficient conditions under which the proposed method exactly represents the population objective, and the boundary example further illustrates the importance of the interior condition for finite-radius exactness.
Critical weaknesses
  • The decision target depends on an untestable, user-chosen bridge class. Regret is minimized over bridges in HB\mathcal H_B, not over causal models. BB sets the null-space component a=B2−∥θmin⁡∥2a=\sqrt{B^2-\|\theta_{\min}\|^2} and hence each contrast-interval width 2a∥PNrδ∥2a\|P_N r_\delta\|, which the observed moments cannot identify, yet BB is never varied (Table 3 varies only interaction terms). The bridge-to-causal gap is large: in the two-world design gbc>0.664g_{bc} > 0.664, with a bridge-contrast endpoint near 1.4441.444 while all compatible causal effects lie in [−0.22,0.78][-0.22,0.78] (Appendix D.5).
  • The finite-sample bounds are uninformative, and exact bridge sets are often empty. EnE_n is linear in Λ\Lambda: with Λ=125\Lambda=125, MY=1M_Y=1, B=2B=2 and n=4,000n=4{,}000, the first two terms of Corollary 8 already exceed 10, far above the reported regrets, and ACIC uses Λ=2415\Lambda=2415. Λ\Lambda comes from population quantities known only because the data are simulated. At ρ=0\rho=0 the empirical bridge set is empty in 25 of 800 matched datasets, which are dropped from Tables 1 and 2 (Appendix E.2), and in 91 of 300 learning-curve datasets.
  • The experiments are small, purpose-built simulations with no established baselines. Every design is built so that admissible bridges reverse the ranking, the policy classes are constant treatment probabilities or two covariate groups with four vertices, and ACIC supplies only covariates. The baselines are variants of the method itself; existing proximal policy learners (Qi et al., 2024; Sverdrup and Cui, 2023), confounding-robust methods and real data are absent. E0E_0 and R0R_0 are the method's own objective, so winning on them is expected, and with a true effect of −1/30-1/30 in the bridge-selection design, part of the causal-loss gain over the always-treat minimum-norm rule comes simply from treating less.
Questions for authors
  1. Which technical steps go beyond Zhang et al. (2023), d'Adamo (2021) and Ben-Michael (2025)? Does the method reduce to an existing one under rectangular restrictions?
  2. How should BB and the bridge features be chosen, and how do the learned policy, excess regret and causal loss change with BB?
  3. What are EnE_n, or the Appendix C.5 bounds, numerically at n=4000n=4000 in each design?
  4. How do Tables 1 and 2 change under a common ρ>0\rho>0 that keeps all datasets? How are ρ\rho and Λ\Lambda chosen when the population moment matrix is unknown?
  5. How does the method compare with existing proximal or confounding-robust policy learners, on richer policy classes or real data?
9 comments
AuthorsOpenReview ↗

Response to Reviewer 4QWY (1/6)

Thank you for recognizing the importance of policy choice under nonunique bridges, the role of joint restrictions, and our theoretical analysis! We address the technical contribution, bridge-class specification, finite-sample guarantees, practical parameter choices, and empirical comparisons below.

[Q1] Technical contributions and the rectangular case

We learn policies from estimated observable moments even when admissible bridges disagree on policy rankings, while preserving their joint restrictions. Under rectangularity, the population criterion recovers existing regret criteria when policy classes and randomization conventions match.

1. Our bounded dual representation connects empirical optimization to population joint-bridge regret without requiring policy rankings to be identified.

  • Zhang et al. (2023) characterize identified functionals, which we apply in Proposition 1. Theorem 3 addresses the remaining decision problem when policy contrasts vary across bridges. In the finite-feature setting, a complete conditional moment representation, positive rank, and a strict interior bridge solution yield uniformly bounded dual coefficients. With a sufficiently large dual radius, these coefficients represent the worst-case value of every relevant contrast exactly at ρ=0\rho=0.
  • d'Adamo (2021, Remark 2) already considers regret within the same policy class. Our convex program retains the common bridge equation and global norm restriction, so adverse contrasts across covariate values must be compatible with a common bridge. It jointly optimizes policy weights and dual coefficients over a finite policy hull.
  • In the cited Ben-Michael (2025) framework, the constraint matrix is known. Here, the bridge matrix must also be estimated and may be rank deficient. Our analysis controls this estimation error along with the other observable moments and transfers these errors to population bridge regret.

Together, Theorem 3 and Corollary 8 give OP(n−1/2)O_P(n^{-1/2}) excess bridge regret under the conditions summarized in W2 below.

2. Rectangularity recovers existing population minimax-regret criteria when the comparator class and randomization conventions match.

To answer your second question directly, let τ(x)\tau(x) be the conditional treatment contrast induced by a bridge. If its set contains every measurable selection from [l(x),u(x)][l(x),u(x)], contrasts can vary independently across covariate values. For treatment-probability policies p,q∈Πp,q\in\Pi,

RΠ(p)=sup⁡q∈ΠE[u(X)(q(X)−p(X))+−l(X)(p(X)−q(X))+],R_{\Pi}(p)=\sup_{q\in\Pi}\mathbb{E}\left[u(X)\bigl(q(X)-p(X)\bigr)_{+}-l(X)\bigl(p(X)-q(X)\bigr)_{+}\right],

Here, XX denotes deployment covariates, l,ul,u are the contrast bounds, and z+=max⁡(z,0)z_+=\max(z,0). The worst case chooses uu where q>pq > p and ll where q<pq < p, and rectangularity permits these choices simultaneously.

For deterministic policies, this is the regret objective within the same policy class discussed by d'Adamo (2021). If the comparator class contains every pointwise optimal policy, it also matches the unrestricted comparator in his main analysis. Unrestricted pointwise randomization for both policies recovers Cui's (2021) rule. In particular, where l(x)<0<u(x)l(x) < 0 < u(x),

p∗(x)=u(x)u(x)−l(x).p^*(x)=\frac{u(x)}{u(x)-l(x)}.

These correspondences concern the population criterion and optimal rule. Equality of finite-sample algorithms also depends on how the unknown quantities are estimated. The revised manuscript states these technical distinctions in Section 2 and the rectangular correspondence in Section 3.2, with the derivation in Appendix A.6.

Thus, the technical contribution connects empirical optimization to the joint population bridge target even when policy rankings remain unidentified.

AuthorsOpenReview ↗

Response to Reviewer 4QWY (2/6)

[W1, Q2] Bridge specification and causal interpretation

Minimizing bridge regret limits worst-case policy loss within a declared bridge class, whose decision impact we quantify by varying BB.

Your suggestion led us to report both the learned policies and their bridge and causal losses.

1. Bridge features and their normalization define the model, and BB limits the remaining bridge variation.

Choose features according to the proxies' causal roles and the assumed bridge structure, fix their normalization, and state the structural or external support for the BB range. In our synthetic bridge-selection design, each treatment arm uses (1,Q,R)/3(1,Q,R)/\sqrt3, where Q,RQ,R are outcome-inducing proxies. The six coefficients jointly satisfy ∣θ∣2≤B\\|\theta\\|_2\le B. The causal bridge constructed in Appendix D.2 is

ha†=12−a30+3Q10+R5h_a^\dagger=\frac12-\frac a{30}+\frac{3Q}{10}+\frac R5

and its coefficient norm under this normalization is

∣θ†∣2=131/60≈1.4776.\\|\theta^\dagger\\|_2=\sqrt{131/60}\approx1.4776.

Thus, the bridge set contains this design's causal bridge for all three values B=1.5,2,2.5B=1.5,2,2.5. This check uses the known generating law. In applications, BB declares the allowed variation, and observed fit or feasibility cannot establish causal-bridge inclusion. Q4 distinguishes an observed-data scale for BB, its modeling interpretation, and sensitivity checks in RHC.

2. Increasing BB lowers treatment probability in this design, but causal loss and bridge regret favor different choices.

We use the same 200 synthetic datasets of size 4,000, changing only the training radius. For evaluation radius bb, let ℓb,ub\ell_b,u_b be the population bridge set's treatment-contrast bounds. For a constant treatment probability p∈[0,1]p\in[0,1],

R0(p;b)=max⁡0,(1−p)ub,−pℓb,R_0(p;b)=\max\\{0,(1-p)u_b,-p\ell_b\\},

E0(p;b)=R0(p;b)−min⁡q∈[0,1]R0(q;b).\mathcal E_0(p;b) = R_0(p;b)-\min_{q\in[0,1]}R_0(q;b).

The two excess-regret columns use the training radius and the common original radius B=2B=2, respectively. Training uses empirical moments, evaluation uses population moments, and each policy is evaluated before averaging. The true effect is −1/30-1/30, so generating-world causal loss is C(p)=p/30C(p)=p/30. Lower loss is preferred for a fixed target.

Table A. Sensitivity to BB in synthetic bridge selection. Values are unscaled means over 200 fitted policies.

Radius BBTreatment ppExcess regret E0↓\mathcal E_0\downarrowExcess regret E0↓\mathcal E_0\downarrowCausal loss C↓C \downarrow
Training BBFixed B=2B=2
1.50.8434650.0064890.0273110.028116
2.00.6818780.0070750.0070750.022729
2.50.6311680.0071860.0173140.021039

Increasing BB from 1.5 to 2.5 lowers mean treatment probability from 0.843465 to 0.631168. Both B=1.5B=1.5 and B=2.5B=2.5 increase excess regret against the common B=2B=2 target. The small excess regrets at each radius's own target concern different objectives. In contrast, causal loss falls as treatment decreases because treatment is harmful in this generating world. Hence, lower causal loss here cannot by itself justify a larger bridge class.

3. Requiring a common bridge excludes incompatible worst-case contrasts, while the causal guarantee retains the bridge-to-causal gap.

Under the stated proxy and bridge-class assumptions, Proposition 4 relates bridge regret R0R_0 to worst-case regret over compatible causal models RcaR_{\mathrm{ca}}:

Rca(π)≤R0(π)≤Rca(π)+gbc,R_{\mathrm{ca}}(\pi) \le R_0(\pi) \le R_{\mathrm{ca}}(\pi)+g_{\mathrm{bc}},

The quantity gbcg_{\mathrm{bc}} is the largest gap between the worst-case policy contrasts allowed by the two classes. Its lower bound of approximately 0.664 in Appendix D.5 shows that observed bridge contrasts can be substantially wider than those of compatible causal models. More data reduce learning error but do not remove this population gap, so bridge optimality need not imply causal optimality.

The decision value is to minimize the largest loss relative to the best policy under each admissible bridge when rankings remain unresolved. Within each policy comparison, the same bridge determines contrasts across covariate values, ruling out incompatible combinations of adverse contrasts. The joint-restriction comparisons in Tables C and D assess the resulting decisions using both bridge regret and generating-world causal loss.

The revised manuscript makes causal-bridge inclusion explicit in Section 3.1 and reports the sensitivity results and their interpretation in Appendix F.1 (Table 5).

Thus, BB is a modeling choice, and observed fit alone cannot verify that the chosen class contains a causal bridge. Even when inclusion holds, the bridge-to-causal gap means that bridge optimality need not imply causal optimality.

AuthorsOpenReview ↗

Response to Reviewer 4QWY (3/6)

[Q3] Finite-sample bounds at the experimental sample sizes

At these sample sizes, the sampling-error bounds in Table B are too large to give an informative numerical bound on excess bridge regret. Under the stated conditions, excess bridge regret still decreases at the root-n rate.

1. The analytic bound grows with the dual radius, giving conservative constants for the experimental settings in Table B.

Using Eq. (22), we set α=0.05\alpha=0.05 and αm=αΓ=0.025\alpha_m=\alpha_\Gamma=0.025. At ρ=0\rho=0, En=en(αm)E_n=e_n(\alpha_m). Outcomes lie in [0,1][0,1], and bridge and test feature norms are at most 1, giving

En≤E‾n=2Λ(1+B)+22,BJn+6[Λ(1+B)+2,B]log⁡(12/α)2n,E_n\le \overline E_n = \frac{2\Lambda(1+B)+2\sqrt2\\,BJ}{\sqrt n} + 6\bigl[\Lambda(1+B)+\sqrt2\\,B\bigr] \sqrt{\frac{\log(12/\alpha)}{2n}},

Here, B,ΛB,\Lambda bound the bridge and dual coefficient norms. The policy hull has J=2J=2 vertices for constant probabilities and J=4J=4 for two covariate groups.

Table B. Analytic upper bounds on EnE_n at n=4,000n=4{,}000.

Experimental design(B,Λ,J)(B,\Lambda,J)Analytic upper bound E‾n\overline E_n on EnE_n
Synthetic bridge selection (all proxy strengths) and criterion comparison(2,125,2)(2,125,2)71.373183
Synthetic and ACIC joint-restriction comparisons(5,2415,4)(\sqrt5,2415,4)1475.348553
ACIC bridge selection and criterion comparison(22,2415,4)(2\sqrt2,2415,4)1745.483093
Two-world learning experiment(3,20,2)(3,20,2)16.027967

2. The sampling contribution to the excess-regret bound is at most twice the listed values, and it decreases at rate n−1/2n^{-1/2} for fixed features, policy class, and norm bounds.

Appendix C.5 now gives the calculation and individual values, complementing the finite-sample analysis in Section 5.1.

Thus, although the current bound remains conservative at these sample sizes, the paper establishes a key learning guarantee under the stated conditions: even with nonunique bridges and unidentified policy rankings, learning from estimated observable moments achieves OP(n−1/2)O_P(n^{-1/2}) excess bridge regret. Sharper numerical bounds at practical sample sizes remain future work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.