acceptodds
Under review as a conference paper at ICLR 2027

ProxMAR: Neural Proximal Causal Inference with Missing Outcome Proxies

Abstract

Proximal causal inference enables causal effect estimation under unmeasured confounding by leveraging proxy variables, but existing neural proximal estimators typically assume that these proxies are fully observed. When outcome-inducing proxies are missing at random (MAR), complete-case analysis can distort both bridge learning and the population averaging used to estimate the causal mean. We propose ProxMAR, a unified framework that addresses proxy missingness at both stages. Rather than filling in missing proxy values, ProxMAR applies conditional-mean imputation, inverse probability weighting, and augmented inverse probability weighting to the estimator-specific functions required for bridge learning and causal-mean evaluation. This avoids modeling the full conditional distribution of potentially high-dimensional proxies. We instantiate the framework with Neural Maximum Moment Restriction (NMMR) and Deep Feature Proxy Variables (DFPV), yielding three missingness-aware variants of each strategy. We establish the observed-data recovery identities underlying these corrections and derive an exact product-remainder identity characterizing the double robustness of augmentation with respect to the missingness propensity and the relevant conditional-mean regression. Under method-specific regularity conditions, we establish consistency of the resulting causal-mean estimators. Numerical studies with low- and high-dimensional proxies demonstrate substantial error reductions relative to complete-case estimation across a range of missingness settings. A semi-synthetic study further shows improved recovery of method-specific full-data estimates.

Primary area:
causal reasoning
Keywords:
Proximal causal learning, Missing at random, Neural network

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Cc6bOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProxMAR extends neural proximal causal inference to settings where outcome-inducing proxies are missing at random. It integrates conditional-mean imputation, inverse probability weighting, and augmented weighting into NMMR and DFPV, addressing missingness in bridge learning and causal-mean evaluation. The paper establishes conditional consistency guarantees and evaluates the resulting estimators on synthetic and semi-synthetic datasets.

Critical strengths

The paper addresses an important practical limitation of proximal estimators through a unified framework that corrects estimator-specific contributions without requiring explicit imputation of high-dimensional proxy values. Its treatment of both bridge learning and population averaging, together with an explicit product-remainder characterization of double robustness, provides a coherent methodological foundation. The experiments demonstrate substantial improvements over complete-case estimation in many settings, particularly at higher missingness rates.

Critical weaknesses
  1. It seems that several consistency assumptions are not connected to the actual implementation. Verifying these conditions or characterizing the potential persistent bias when this assumption fails would be helpful for better transparency. Specifically:

    1. C.5: The paper provides sufficient conditions linking vanishing moment risk to causal-mean consistency, but does not establish that these conditions hold for the implemented neural class.
    2. C.6 and C.9: The required uniform control over candidate bridges is not established for the implemented kernel regressions. In particular, the fixed bandwidth grid may leave persistent smoothing bias.
    3. C.4: The theory assumes that the candidate neural class contains an exactly valid bridge, without verifying this for the implemented architecture.
    4. D.6: The theory requires learning errors to vanish sufficiently quickly relative to the inverse-stability factor. No concrete regime is established in which network capacity, regularization, and nuisance estimation jointly satisfy this requirement.
  2. For the high-dimensional experiment, although WW has 100–300 coordinates, it is generated from only two latent factors through a linear Gaussian measurement model. The true bridge is additive and linear in WW, and nuisance regressions always condition on just four variables, L+=(A,Z1,Z2,Y)L^+=(A,Z_1,Z_2,Y). All three main experimental designs also omit observed covariates XX. This design provides limited evidence for complex settings. Maybe the authors can try varying latent dimension, proxy informativeness, nonlinear measurement structure, and the dimension of XX to better test the scalability claim.

Questions for authors
  1. Appendix E.1.2 acknowledges that the Demand benchmark’s observation probabilities are not uniformly bounded away from zero, whereas Assumption 2 requires e(L+)≥ϵ>0e(L^+)\geq\epsilon>0. Could the authors clarify which guarantees remain applicable to this benchmark and whether consistency can be established under weaker overlap or inverse-probability moment conditions?

  2. For the continuous-treatment experiments, what properties of the neural class and DGP ensure Assumption D.7(ii), and how can the growth of Ca,nC_{a,n} be controlled?

  3. Tables 1–5 compare ProxMAR primarily with complete-case estimation and simple mean imputation within DFPV and NMMR. Could the authors include a conditional distribution-based or multiple-imputation baseline for WW given L+=(A,Z,X,Y)L^+=(A,Z,X,Y), to assess the claimed advantage of recovering estimator-specific functions directly in terms of accuracy and computational cost?

  4. For table1-5, it seems that AIPW is frequently worse than IMP or IPW, sometimes substantially so, particularly for DFPV. What could be the explanation? Under what practical conditions does its additional robustness become beneficial?

  5. Could the authors explain the settings where complete-case estimation outperforms all corrections? For example, Demand NMMR at 30% missingness (Table 1) and Demand DFPV at (n=1000) (Table 2)

  6. Also in tables 2 and 5 it seems that the corrected estimators worsen as sample size increases. What explains this pattern?

8 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1.1

Critical Weakness 1.1: Assumption C.5 -- Target-mean Separation

We have revised the NMMR theory to accommodate growing neural bridge classes and clarified the identification, approximation, complexity, and nuisance recovery conditions required for consistency.

Assumption C.5 links vanishing population NMMR risk to causal-mean consistency. For the neural bridge class Hn={hθ:θ∈Θn}H_n = \{h_\theta: \theta \in \Theta_n\}, define

ba(θ)=E[hθ(W,a,X)]b_a(\theta) = \mathbb{E}[h_\theta(W,a,X)]

ca,n(ε)=inf⁡{Rκ(θ):θ∈Θn,∣ba(θ)−β(a)∣≥ε}c_{a,n}(\varepsilon) = \inf \{R_\kappa(\theta): \theta \in \Theta_n, |b_a(\theta)-\beta(a)| \geq \varepsilon\}.

The revised condition requires lim inf⁡n→∞ca,n(ε)>0\liminf_{n\to\infty} c_{a,n}(\varepsilon) > 0 for every ε>0\varepsilon > 0.

Sufficient conditions. Appendix C.2 establishes three conditions for this uniform separation:

(i) Identification. Every zero-risk function in the common function-class closure KK must identify the target causal mean: Rκ(h)=0R_\kappa(h)=0 implies E[h(W,a,X)]=β(a)\mathbb{E}[h(W,a,X)]=\beta(a).

This follows from proximal identification when zero-risk candidates satisfy the required outcome-bridge and integrability conditions. For continuous treatment, bridge validity at the specified treatment value and the pointwise conditions in Appendix A.1 are additionally needed. Bridge uniqueness is not required.

(ii) Continuity. Both the population risk Rκ(h)R_\kappa(h) and target mean ba(h)b_a(h) must be continuous on KK in the same topology. A weighted supremum norm provides one sufficient setting under appropriate moment conditions and a bounded NMMR kernel. For a fixed finite ReLU network with bounded parameters, a linear-growth envelope and finite input moments can provide the required continuity.

(iii) Common compactness. The closure KK of the union of the neural classes HnH_n must be compact in the chosen function topology. For fixed architectures, bounded parameter sets and continuity of the network imply compactness of the corresponding function class. For growing classes, however, compactness of each parameter space Θn\Theta_n alone is insufficient; uniform control across widths is required. On compact input domains, uniform boundedness and equicontinuity provide sufficient conditions through the Arzelà–Ascoli theorem.

Under these three properties, the set of functions in KK whose target means differ from β(a)\beta(a) by at least ε\varepsilon is compact and contains no zero-risk function. By continuity, the population risk attains a strictly positive minimum on this set whenever it is nonempty. This minimum uniformly bounds ca,n(ε)c_{a,n}(\varepsilon) away from zero, establishing C.5.

Connection to implementation. Prescribed parameter-box constraints can be enforced by projection after each optimizer update (Appendix C.1.3 and Algorithm 1). Such constraints make each finite-dimensional parameter space compact but do not themselves guarantee common compactness for growing-width networks.

The reported benchmarks use finite, fixed-width architectures. We do not claim that their numerical configurations verify all the above sufficient conditions. Instead, the revised theory characterizes conditions under which neural classes satisfy C.5, while the experiments assess finite-sample performance. Approximation requirements for growing classes are addressed separately under Assumption C.4.

Revisions. Section 3.1 introduces sample-size-dependent neural bridge classes; Section 4 states the revised target-mean separation condition. Appendix C.2 establishes sufficient conditions based on common compactness, continuity, and identification of zero-risk bridges. Appendix C.1.3 and Algorithm 1 describe projection onto prescribed parameter constraints and distinguish these constraints from the approximate-minimization requirement.

Critical Weaknesses 1.2 and 1.3: Uniform Nuisance Control and Neural Bridge Specification

1.2. Assumptions C.6/C.9: Uniform nuisance control

**Bandwidth selection.**Fixed bandwidth grids may retain smoothing bias; cross-validation (CV) over a fixed grid alone cannot establish C.6 or C.9. Original Demand/Gaussian benchmarks retain fixed bandwidths for finite-sample comparison.

The nonlinear high-dimensional study (Appendix F.2) and MNAR sensitivity analysis (Appendix I) use five-fold CV over 30 random hyperparameter configurations (Appendix F, Table 10). For IMP/AIPW, candidate bandwidths depend on sample size mm and conditioning dimension dd: Bm,d={2j/2hROT:j=−2,−1,0,1,2},B_{m,d} = \{2^{j/2} h_{\mathrm{ROT}} : j = -2,-1,0,1,2\}, with hROT=dm−1/(d+4)h_{\mathrm{ROT}} = \sqrt{d} m^{-1/(d+4)}.

For fixed dd, this grid shrinks with mm and meets the usual bandwidth/effective-sample-size scaling, but this alone does not establish uniform nuisance consistency.

C.6 uniformity. In NMMR, the Nadaraya-Watson (NW) residual smoother uses shared nonnegative weights summing to one. Bandwidth selection uses initial training residuals; the weights stay fixed during bridge optimization (Appendix C.1.3). For bridges h,gh,g in the common compact class KK, these weights imply ∣m^rh(ℓ+)−m^rg(ℓ+)∣≤∥h−g∥vv^(ℓ+),|\widehat m_{r_h}(\ell^+) - \widehat m_{r_g}(\ell^+)| \leq \|h-g\|_v \widehat v(\ell^+),

where vhat is the NW regression of the envelope v(W,A,X) with those weights. If NW is L2-consistent for fixed residual responses and fitted envelope regressions are uniformly L2-bounded in probability, finite covers give uniform consistency over K; cross-fitting transfers this to C.6's held-out empirical errors (Appendix C.4.1).

C.9's two requirements. First, under a fitted propensity floor c>0c>0, the AIPW product remainder is bounded by

c−1∥eˉ−e∥L2sup⁡θ∥m^rθ−mrθ∥L2.c^{-1}\lVert\bar e-e\rVert_{L_2}\sup_\theta\lVert\widehat m_{r_\theta}-m_{r_\theta}\rVert_{L_2}.

It vanishes if either nuisance error vanishes and the other remains bounded.

Second, C.9 separately requires a centered uniform empirical-process condition. Appendix C.4.3 gives sufficient conditions through cross-fitting, a bounded kernel, suitable envelopes and finite covers; the oracle-class complexity condition is separate. These arguments do not claim that the original fixed numerical configurations satisfy all asymptotic assumptions.

Revisions for 1.2. Appendix E separates fixed finite-sample bandwidths from asymptotic guarantees; Appendix F/Table 10 lists data-driven candidates for F.2/I; Appendices C.4.1 and C.4.3 give sufficient uniform-control arguments.

1.3. Assumption C.4: Approximate neural bridge specification

We replaced exact bridge inclusion in the implemented network class with approximate specification for growing classes, motivated by neural approximation theory (e.g., Hornik, 1991; Schmidt-Hieber, 2020; Farrell et al., 2021). For a valid bridge h0h_0 and bounded MMR kernel,

Rκ(θ)≤Cκ∥hθ−h0∥L2(PW,A,X)2.R_\kappa(\theta) \leq C_\kappa \lVert h_\theta-h_0\rVert_{L_2(P_{W,A,X})}^2.

Thus L2L_2 approximation can drive risk to zero even if h0h_0 is never exactly in HnH_n. Revised C.4 asks for a deterministic comparison sequence θn∗\theta_n^* in Θn\Theta_n with

Rκ(θn∗)+λn∥θn∗∥22→0.R_\kappa(\theta_n^{\ast}) + \lambda_n \lVert\theta_n^{\ast}\rVert_2^2 \to 0.

Both approximation and the penalty matter: increasing width alone is insufficient. Appendix C.3, Eq. (61), gives for each outer fold kk and correction jj:

Rκ(θ^j(−k))≤2Γn,k,j+Rκ(θn∗)+λn∥θn∗∥22+op(1),R_\kappa(\widehat\theta_j^{(-k)}) \leq 2\Gamma_{n,k,j} + R_\kappa(\theta_n^{\ast}) + \lambda_n \lVert\theta_n^{\ast}\rVert_2^2 + o_p(1),

where Γ\Gamma is the uniform corrected-criterion deviation and the last term includes approximate-minimization error. Together with uniform estimation/optimization, C.4 yields vanishing risk; C.5 and held-out evaluation conditions then imply causal-mean consistency.

Connection to implementation. The original Demand NMMR uses a fixed three-hidden-layer width-80 network and penalty 3×10−63 \times 10^{-6} across sample sizes; we retain it but do not claim it satisfies the asymptotic conditions.

Since its valid bridge hDh_D is known, fitting the same architecture to hDh_D on simulated full data and measuring held-out L2L_2 error could upper-bound its population moment risk, although we did not run this diagnostic.

For growing networks with pnp_n parameters bounded coordinatewise by BB, λnpnB2→0\lambda_n p_n B^2 \to 0 suffices for the penalty contribution; approximation, uniform estimation and optimization still need control.

F.2 and I instead tune 30 randomly sampled configurations by five-fold CV, including architecture, learning rates and regularization; finite-candidate tuning also does not prove C.4.

Revisions for 1.3. Section 3.1/Appendix C.2 introduce growing classes and approximate specification; Theorem 1/Appendix C.3 incorporate approximation and penalty errors. Appendix E documents fixed configurations, while Appendix F/Table 10 reports the expanded tuning protocol.

Critical Weakness 1.4 & Question 3.2: Inverse stability and fixed-treatment evaluation

The original Assumptions D.6 and D.7(ii) are strong and are not guaranteed by a neural architecture or hyperparameter tuning. In particular, controlling the inverse-stability factor τn\tau_n and the fixed-treatment transfer factor Ca,nC_{a,n} requires additional restrictions on the candidate bridge class.

Alternative local dual-bridge proof. We revised Theorem 2 to establish an alternative route to DFPV causal-mean consistency without D.6 or D.7(ii). The two-stage analysis bounds the projected error

εn,j≥max⁡k∥E[h^j(−k)(W,A,X)−Y∣L]∥L2(PL)=op(1)\varepsilon_{n,j} \geq \max_k \lVert \mathbb{E}[\widehat h_j^{(-k)}(W,A,X)-Y \mid L] \rVert_{L_2(P_L)} = o_p(1)

Appendix D.4 separates neural approximation, nuisance estimation, empirical-risk deviations, regularization, and optimization errors.

Assumption D.8 instead uses a local dual treatment bridge q(t,Z,X)q(t,Z,X) satisfying, for almost every tt near the target aa,

E[q(t,Z,X)∣A=t,W,X]=dPW,XdPW,X∣A=t\mathbb{E}[q(t,Z,X) \mid A=t,W,X] = \frac{dP_{W,X}}{dP_{W,X\mid A=t}}

and E[q(t,Z,X)2∣A=t]≤Cq,a\mathbb{E}[q(t,Z,X)^2 \mid A=t] \leq C_{q,a}. This directly controls causal-mean error through the projected bridge error rather than strong convergence of the entire outcome bridge.

Continuous treatment and pointwise consistency. Convergence integrated over the factual treatment distribution does not alone imply consistency at a fixed continuous treatment value. D.8 therefore also requires local treatment mass PA(∣A−a∣≤δ)≥caδP_A(|A-a|\leq\delta) \geq c_a\delta and local ss-Hölder continuity of the fitted causal-mean errors, with factor MnM_n. Theorem D.1 (Eq. 106) gives

∣β^j(a)−β(a)∣≤Sn,j(a)+Bαn,j+Cq,acaδεn,j+Mnδs|\widehat\beta_j(a)-\beta(a)| \leq S_{n,j}(a) + B\alpha_{n,j} + \sqrt{\frac{C_{q,a}}{c_a\delta}}\varepsilon_{n,j} + M_n\delta^s

where Sn,j(a)=Op(nev−1/2)S_{n,j}(a) = O_p(n_{\mathrm{ev}}^{-1/2}) and αn,j=op(1)\alpha_{n,j} = o_p(1) controls Stage 1 and evaluation errors. If Mn=Op(1)M_n=O_p(1), any vanishing projected error suffices for this local transfer step. More generally, for εn,j=Op(rn)\varepsilon_{n,j}=O_p(r_n) and Mn=Op(mn)M_n=O_p(m_n), the sufficient rate condition is mnrn2s→0m_n r_n^{2s} \to 0.

Appendix D.3.4 bounds the neural mean-curve Hölder factor using the coefficient norm, treatment-feature regularity, and proxy/covariate-feature moments; controlling treatment-network layer spectral norms is sufficient for the Lipschitz contribution.

Comparison of sufficient conditions. We retain D.6 and D.7(ii) as an alternative. For projected error Op(rn)O_p(r_n), that route requires max⁡1,Ca,nτnrn→0\max\\{1,C_{a,n}\\}\tau_n r_n \to 0; the local dual route instead requires the above Hölder-rate condition. The latter avoids these two inverse-stability/transfer factors but introduces substantive assumptions on dual-bridge existence, integrability, local treatment mass, and regularity. Neither route universally dominates, and neither is guaranteed by the implemented architectures.

Revisions. Section 4 (Theorem 2) states both routes; Appendix D.3.4 gives the dual-bridge, local mass, neural Hölder and rate conditions; Appendix D.4 proves the fixed-treatment bound and consistency.

Reviewer_33H7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProxMAR, a neural proximal causal inference framework for estimating average dose-response functions when outcome proxies are missing at random (MAR). Instead of imputing the missing proxies directly, it applies three missing-data corrections (IMP, IPW, and AIPW) to the functions used for bridge learning and causal-mean estimation. The framework is instantiated with two neural proximal methods, NMMR and DFPV. The authors prove recovery identities, the double robustness of AIPW, and consistency, and report good results on synthetic and semi-synthetic data.

Critical strengths
  • Solid theory: the paper gives clear proofs of the recovery identities, the double robustness of AIPW, and the consistency of the estimators.
  • Unified and flexible: it combines three missing-data corrections (IMP, IPW, AIPW) with two different neural proximal methods (NMMR and DFPV) in a single framework.
Critical weaknesses
  • Strong dependence on MAR: the method requires the missing-at-random assumption. If the data are missing not at random (MNAR), performance may degrade, and the paper does not provide a sensitivity analysis for this case. The authors acknowledge this limitation in Section 6 but do not address it.

  • Nuisance estimation in high dimensions is not actually tested. The conditional means and observation probabilities are estimated with Nadaraya-Watson kernel regression on the complete cases, conditioning on all fully observed variables (A, Z, X, Y). Kernel smoothing of this kind is known to suffer from the curse of dimensionality. The "high-dimensional" study in Section 5.2 and Appendix E.2.5 only increases the dimension of the missing proxy W, which has a two-dimensional latent structure by construction, while the dimension of the conditioning set stays fixed. It therefore does not show how the nuisance estimators behave when Z or X is high-dimensional, which is the case that matters for the proposed corrections.

  • The missing-data baselines are too weak. The proposed corrections are only compared with complete-case analysis and mean imputation. Multiple imputation (e.g., chained equations), which is the standard practical approach under MAR, is not included, so it is unclear whether the proposed corrections improve over what a practitioner would already do.

  • There is no evaluation on real data with naturally occurring missingness. The RHC study is semi-synthetic: MAR missingness is imposed artificially on W, and the evaluation measures recovery of each method's own full-data estimate rather than a causal ground truth. In a real application the missingness mechanism is unknown, MAR cannot be verified, and the full-data estimate is unavailable. The paper offers no guidance for this setting: how to assess whether MAR is plausible, how to choose among IMP, IPW, and AIPW, or how to diagnose failure of the nuisance models.

Questions for authors
  1. Could the authors add a multiple-imputation baseline (e.g., MICE followed by NMMR or DFPV) in the Demand and high-dimensional experiments?

  2. How do the authors recommend applying ProxMAR to a dataset where the missingness in W occurs naturally and MAR cannot be checked? In particular, which variant (IMP, IPW, or AIPW) should a practitioner use by default, and are there diagnostics for the nuisance models?

  3. Could the authors report an experiment where the dimension of the conditioning set (Z or X) grows, to test the kernel-based nuisance estimators directly?

5 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1

Critical Weakness 1: Dependence on MAR and MNAR Sensitivity

MAR cannot generally be verified from observed data, and our consistency guarantees do not automatically extend to MNAR. We have clarified the role of MAR and added both a recovery-bias analysis and an MNAR sensitivity experiment.

1. Why study MAR?

MAR has been central to missing-data methodology since Rubin (1976) and remains widely used (Little & Rubin, 2019; Seaman et al., 2013; Little, 2021; Delaigle et al., 2020). Molenberghs et al. (2008) show that an MNAR model can have an MAR counterpart with the same observed-data fit; distinguishing the two generally requires assumptions beyond observed data.

MAR is nevertheless a meaningful working assumption when the collection process and recorded predictors of missingness support it. The National Research Council (2010) and White et al. (2011) advocate analyses under scientifically plausible missing-data assumptions, accompanied by sensitivity analyses. Recent causal research also continues to develop MAR-based methods (Gao et al., 2022; Kallus & Mao, 2025; Sundberg et al., 2026; Pryce et al., 2025).

2. ProxMAR's contribution under MAR.

Missing outcome proxies affect both neural proximal bridge learning and final causal-mean averaging. NMMR learns the bridge through conditional moments of residuals; DFPV first learns conditional proxy features and then regresses the outcome.

ProxMAR integrates IMP, IPW, and AIPW into these estimator-specific steps, directly recovering the required functions without modeling the full conditional distribution of potentially high-dimensional WW. We derive the observed-data recovery identities and method-specific consistency results.

3. Recovery bias under MNAR.

Let e0(L+)=Pr⁡(δW=1∣L+)e_0(L^+) = \Pr(\delta_W = 1 \mid L^+) and ms,d(L+)=E[s∣L+,δW=d]m_{s,d}(L^+) = \mathbb{E}[s \mid L^+, \delta_W = d] for d∈{0,1}d \in \{0,1\}.

When the MAR-based transformations use their correct observed-data nuisance targets, all three have the same conditional recovery bias:

E[Tjobs(s)∣L+]−E[s∣L+]=(1−e0(L+))[ms,1(L+)−ms,0(L+)],j∈{IMP,IPW,AIPW}.\mathbb{E}[T_j^{\mathrm{obs}}(s)\mid L^+]-\mathbb{E}[s\mid L^+]=(1-e_0(L^+))[m_{s,1}(L^+)-m_{s,0}(L^+)],\quad j\in\{\mathrm{IMP},\mathrm{IPW},\mathrm{AIPW}\}.

Thus, differences between observed and missing proxy groups can bias the recovered bridge moments (NMMR), conditional proxy features (DFPV), and ultimately causal-mean estimates. The equal conditional recovery bias does not imply identical fitted estimators or finite-sample behavior.

4. Numerical MNAR sensitivity.

We retain the Gaussian design's complete-data DGP and causal target (Appendix E.2.1) while introducing direct dependence on WW through

eγ(W,L+)=expit⁡(cr,γ+1.6A∘+0.8Z1∘−0.8Z2∘+1.2Y∘+γW∘)e_\gamma(W,L^+)=\operatorname{expit}(c_{r,\gamma}+1.6A^\circ+0.8Z_1^\circ-0.8Z_2^\circ+1.2Y^\circ+\gamma W^\circ),

where W∘W^\circ is the standardized bridge-relevant projection cW⊤Wc_W^\top W.

We fix n=5,000n = 5{,}000, proxy dimension pW=200p_W = 200, and a target missingness rate of 50%50\%; γ∈{0,0.5,1,1.5,2}\gamma \in \{0, 0.5, 1, 1.5, 2\}. The γ=0\gamma = 0 case serves as the MAR reference. Unlike the original Gaussian benchmark, these data-generating observation probabilities are not truncated.

Over 20 replications, mean CMAE increases with γ\gamma for IMP and AIPW in both estimator families, while IPW is non-monotone. From γ=0\gamma=0 to γ=2\gamma=2, AIPW CMAE rises from 0.414 to 0.514 for DFPV and from 0.516 to 0.557 for NMMR. All three corrections outperform their corresponding complete-case estimators over the reported range (Appendix I, Fig. 9 and Table 17).

These findings characterize sensitivity to this particular MNAR mechanism; they do not establish identification or consistency under general MNAR.

Revisions. The Introduction clarifies the motivation, relevance, and limitations of MAR. Appendix B.2 derives the MNAR recovery-bias identity (Lemma B.2) and its implications for NMMR and DFPV. Appendix I specifies the sensitivity experiment, distinguishes its untruncated mechanism from the original clipped Gaussian benchmark, and reports Fig. 9 and Table 17. Section 6 clarifies the limits of the MNAR analysis.

AuthorsOpenReview ↗

Response to Critical weakness 2/Question 3

Critical Weakness 2 & Question 2.3: High-dimensional nuisance estimation

We agree that the original Gaussian proxy experiment increases the dimension of WW while keeping the fully observed conditioning set at dimension four. We clarify that Nadaraya–Watson (NW) regression estimates the conditional-mean nuisances, whereas observation propensities are estimated separately using neural-network classifiers. The concern regarding high-dimensional conditional-mean estimation is nevertheless important.

Additional nonlinear high-dimensional experiment.

We added a nonlinear experiment (Appendix F.2) with n=5,000n = 5{,}000, four latent confounders, and 20-dimensional outcome proxies. Across seven configurations of treatment-proxy dimension pZp_Z and observed-covariate dimension pXp_X, the conditioning dimension d=dim⁡(L+)=pZ+pX+2d = \dim(L^+) = p_Z + p_X + 2 varies from 32 to 62.

For IMP and AIPW, we compare NW with direct neural-network (NN) regression for the conditional-mean nuisances. The following table reports CMAE from one replication per configuration; lower values are better.

pZp_ZpXp_XddDFPV-IMP NWDFPV-IMP NNDFPV-AIPW NWDFPV-AIPW NN
530370.6830.2230.3970.334
1020320.5990.1940.3200.177
1030420.5510.3300.3450.452
2020420.7870.7140.4840.343
2030520.5930.3610.2650.302
3030620.5390.4420.1920.207
3210440.7060.7640.3230.502
pZp_ZpXp_XddNMMR-IMP NWNMMR-IMP NNNMMR-AIPW NWNMMR-AIPW NN
530370.8070.9470.8580.557
1020320.9591.0821.0130.823
1030420.7950.9340.9730.484
2020420.8971.0671.0000.678
2030520.7750.9060.8620.501
3030620.7700.8870.8420.412
3210441.0361.1350.9240.696

NN achieves lower CMAE for DFPV-IMP in six of seven settings and for NMMR-AIPW in all seven. Conversely, NW performs better for NMMR-IMP in all seven, while DFPV-AIPW exhibits mixed results.

These findings demonstrate estimator- and correction-dependent finite-sample performance rather than uniform superiority of either nuisance regression approach. They measure downstream causal-mean error, not direct nuisance prediction error.

The comparisons use one replication per setting and 30 randomly sampled hyperparameter configurations selected by five-fold cross-validation. They therefore do not establish general high-dimensional consistency or a universal performance ranking.

Revisions. Section 5.2 distinguishes high-dimensional missing proxies from high-dimensional nuisance conditioning. Appendix F.2.1 specifies the nonlinear DGP and seven dimension configurations; F.2.2 describes the MAR mechanism and NW/NN comparison; F.2.3 reports the full results (Tables 12–13). Appendix F and Table 10 detail the hyperparameter search spaces and selection procedure.

AuthorsOpenReview ↗

Response to Critical Weakness 3/Question 1

Critical Weakness 3 & Question 2.1: Missing-data baselines

We added multiple imputation by chained equations with predictive mean matching (MICE-PMM), followed by DFPV or NMMR, to the original Demand and Gaussian benchmarks (Appendix E), and further evaluated it in two new nonlinear proxy designs (Appendix F).

1. Accuracy comparisons.

The following table summarizes representative results at n=5,000n = 5{,}000. Entries are mean CMAE (Monte Carlo SE), except for F.2, which reports one replication. Lower CMAE is better.

DesignEstimatorMICE-PMMProxMARCorrectionReps.
DemandDFPV6.058 (0.047)4.897 (0.052)IPW20/100
DemandNMMR3.922 (0.049)3.716 (0.028)IPW20/100
GaussianDFPV0.320 (0.014)0.359 (0.007)IMP20/50
GaussianNMMR0.140 (0.009)0.188 (0.005)IPW20/50
Second moment (F.1)DFPV0.1024 (0.0071)0.0360 (0.0027)IMP56/56
Second moment (F.1)NMMR0.2030 (0.0038)0.0452 (0.0035)IMP56/56
Nonlinear (F.2)DFPV0.7700.330IMP-NN1/1
Nonlinear (F.2)NMMR0.9400.484AIPW-NN1/1

"Reps." gives the numbers of MICE-PMM and ProxMAR replications. The original benchmarks use separate Monte Carlo batches, whereas F.1 uses 56 paired replications. The Gaussian comparison fixes proxy dimension 200 and 50%50\% missingness; F.2 illustrates (pZ,pX)=(10,30)(p_Z, p_X) = (10, 30) with proxy dimension 20.

2. Why does MICE-PMM perform well in the original benchmarks?

Its favorable performance may reflect relatively simple proxy structures. In Demand, WW is scalar and strongly associated with (A,Z,Y)(A, Z, Y); predictive mean matching uses observed donor values, potentially preserving nonlinear features. In the Gaussian design, although WW is high-dimensional, its latent signal is two-dimensional, and the bridge depends linearly on a projection of WW predictable from the observed variables.

3. Why do the nonlinear designs differ?

In F.1, E(W∣L+)=0\mathbb{E}(W \mid L^+) = 0, but E(W2∣L+)\mathbb{E}(W^2 \mid L^+) varies with L+L^+, and the bridge depends on WW squared. All three ProxMAR corrections outperform both MICE-PMM and unconditional mean imputation for DFPV and NMMR across 56 paired replications (Table 11).

In F.2, the bridge depends on a squared projection of WW whose conditional mean is zero but whose conditional second moment varies with L+L^+. DFPV-IMP with NN regression and NMMR-AIPW with NN regression outperform MICE-PMM in all seven high-dimensional configurations (Tables 12–13). These single-replication results are descriptive.

Although PMM uses observed donor values, matching on predicted conditional means does not explicitly ensure recovery of the higher-order moments or joint dependence required by these bridges. More flexible imputation models may perform better, and imputation methods may be sensitive to misspecification under complex nonlinear or high-dimensional proxy dependence.

ProxMAR instead directly recovers estimator-specific quantities without modeling the full conditional distribution of WW given L+L^+. It still requires accurate nuisance estimation. Thus, these comparisons do not establish uniform superiority over multiple imputation.

Revisions. Appendices E.1–E.2 add MICE-PMM comparisons to the Demand and Gaussian benchmarks (Tables 1–2 and 6–8). Appendix E.1.5 reports computational costs (Table 3). Section 5.1 and Appendix F.1 introduce the second-moment design (Table 11), while Section 5.2 and Appendix F.2 present the nonlinear high-dimensional experiment (Tables 12–13).

Reviewer_JJLbOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper addresses proxy causal learning (PCL) in missing outcome proxy variable setting. PCL identifies and learns causal functions under unobserved confounding setting by utilizing informative proxy variables. This work extends the existing PCL algorithms to estimate dose-response in missing at random outcome proxy case. Three methods are adopted to prior PCL algorithms. The augmented inverse probability weighting yields a product remainder characterizing double robustness. The methods are tested in synthetic and semi-synthetic studies.

Critical strengths

Overall, the paper addresses an important problem in proxy causal learning.

  • The paper addresses a practically relevant limitation of existing PCL methods: missing outcome proxy variables. This is practically relevant as missing data is a common problem in machine learning and causal inference.
  • It develops estimators with consistency guarantees under stated assumptions and incorporates missing-data corrections into NMMR (Kompa et al.) and DFPV (Xu et al.). These extensions broaden the settings in which existing PCL algorithms can be applied.
  • The synthetic and semi-synthetic experiments show that the proposed methods often improve upon the corresponding naive complete-case implementations of NMMR and DFPV. These results provide empirical support for explicitly accounting for missingness.
Critical weaknesses
  • As it is written, it is not clear to me how the product-remainder identity contributes to the proposed methods. I can locate the derivations in the appendix; yet I believe not sufficient details are discussed in the main paper. This is an important property of the method but it is not clearly presented in the main body.
  • For the NMMR variant of the method, the authors choose to use V-statistics based loss function which is a biased estimator of the risk. The appendix also briefly discusses the U-statistics variant but it is quite brief in my opinion. Also, judging from Figure 2 of Kompa et al., it looks like U-statictic variant is much more stable in demand experiment. Given that the paper also presents demand experiment, I am curios why the V-statistic variant is chosen.
  • Assumption D.6 states that d2(hη,H0)≤τ∥E[Tj(hη)∣L]−E[Y∣L]∥d_2(h_{\eta}, \mathcal{H}_0) \leq \tau \lVert \mathbb{E}[T_j(h_{\eta}) \mid L] - \mathbb{E}[Y \mid L] \rVert. This is a strong assumption because it bounds the strong bridge consistency by τn\tau_n times the projected bridge consistency. (This is I believe the analogous assumption of τa\tau_a in Xu et al. right after their Assumption 3) In the ill-posed case, the convergence on the left hand side can be very slow which requires τn\tau_n to increase as the candidate class expands. One way to avoid this is to use the dual treatment bridge function and characterize the causal function consistency in terms of this dual function and the projected outcome bridge consistency; see Theorem 4.1 in (arXiv 1810.00283), and Remark H.10 and Theorem H.13 in (arxiv 2605.09514). I understand that this comes with different set of assumptions and a substantial proof change; yet I am curious if you can quantify this balance for the neural classes.
  • The main text does not direct reader to the proof of Proposition 1.
Questions for authors
  • Related to the second item above, I am wondering why authors have chosed V-statistic method of NMMR?
  • Related to the third item above, can you justify Assumption D.6?
  • For the high dimensional proxy experiment with DFPV, it is interesting to see that AIPW method's error increases as the missing rate increases whereas the IMP and IPW variants stay stable. Do authors have any intiution as to why?
5 comments
AuthorsOpenReview ↗

Response to Critical weakness 1, and weakness 4

Critical Weakness 1: Contribution of the AIPW Product-Remainder Identity

1. Product-remainder identity and methodological motivation.

Lemma B.1 (Eq. 26) establishes that the conditional recovery bias of AIPW is governed by the product of the propensity and conditional-mean specification/estimation errors, up to an inverse-probability factor. Unlike IMP and IPW, whose recovery biases depend on the respective nuisance errors individually, AIPW requires only a vanishing product remainder under suitable regularity conditions. This can hold even when only one nuisance estimator is consistent, provided the other estimation error is sufficiently controlled.

2. Contribution to ProxMAR constructions and consistency theory.

This identity directly motivates our AIPW constructions. For NMMR, the AIPW transformation is applied to the bridge residual during training and to the bridge evaluation during causal-mean estimation. For DFPV, it is applied to the proxy features in Stage 1. The corresponding product-remainder conditions control nuisance-induced bias in our consistency analyses, together with the remaining bridge-learning and empirical-process conditions.

Revisions. We added a discussion immediately after Proposition 1 in Section 2.2 explaining the exact product-remainder property and its implication for asymptotic recovery bias. Sections 3.1 and 3.2 now clarify its role in NMMR bridge learning and causal-mean evaluation, and DFPV Stage 1 proxy-feature learning, respectively. Section 4 further distinguishes the nuisance consistency requirements of IMP, IPW, and AIPW.

Critical Weakness 4: Proof of Proposition 1

Proposition 1 follows from Lemma B.1 in Appendix B.1: setting e~=e\widetilde e=e and m~s=ms\widetilde m_s=m_s in Eqs. (24)–(26) makes all conditional remainders zero; iterated expectations give the remaining identities. We added this explicit proof pointer immediately after Proposition 1 in Sec. 2.2.

AuthorsOpenReview ↗

Response to Critical weakness 2/Question 1

Critical Weakness 2 & Question 1.1: Why NMMR-V rather than NMMR-U?

1. Motivation. Let r^j,i(θ)=T^j,i(rθ,i)\hat r_{j,i}(\theta)=\hat T_{j,i}(r_{\theta,i}) be the corrected residual for jj = IMP, IPW, or AIPW.

Let r^j(θ)\hat r_j(\theta) collect these residuals over NN training observations, and let Kiℓ=κ(Li,Lℓ)K_{i\ell}=\kappa(L_i,L_\ell). The unpenalized objectives are

R^V,j(θ)=r^j(θ)⊤Kr^j(θ)N2=∥1N∑i=1Nr^j,i(θ)κ(Li,⋅)∥Hκ2.\hat R_{V,j}(\theta)=\frac{\hat r_j(\theta)^\top K\hat r_j(\theta)}{N^2}=\left\lVert\frac{1}{N}\sum_{i=1}^N\hat r_{j,i}(\theta)\kappa(L_i,\cdot)\right\rVert_{\mathcal H_\kappa}^2.

R^U,j(θ)=NN−1R^V,j(θ)−1N(N−1)∑i=1NKiir^j,i(θ)2.\hat R_{U,j}(\theta)=\frac{N}{N-1}\hat R_{V,j}(\theta)-\frac{1}{N(N-1)}\sum_{i=1}^N K_{ii}\hat r_{j,i}(\theta)^2.

The U criterion involves K−diag⁡(K)K-\operatorname{diag}(K), which is indefinite whenever nonzero. Its parameter-dependent diagonal subtraction may decrease U even when the squared empirical kernel moment increases. Both objectives include the same penalty λn∥θ∥22\lambda_n\lVert\theta\rVert_2^2. Regularization may mitigate but need not eliminate this instability. In contrast, the unpenalized V loss directly measures the squared empirical RKHS moment norm.

2. Demand comparisons. At n=1000n=1000 and 50%50\% missingness, we ran 100 replications per method. The results are mean CMAE (Monte Carlo SE).

VersionNMMR-UNMMR-V
Full data4.214 (0.041)3.955 (0.036)
Complete case5.583 (0.155)7.602 (0.150)
Unconditional mean (UMI)23.022 (16.415)7.652 (0.150)
IMP3.994 (0.040)3.811 (0.036)
IPW5.656 (0.990)3.069 (0.059)
AIPW106233.970 (60127.313)3.919 (0.042)

To investigate the observed U instability, we generated another independent Demand sample (n=400n=400, 50%50\% missingness). U and V were trained on the same 320 observations (80%80\%), with identical architectures, initialization, and training/validation settings. The following reports unpenalized criteria and maximum absolute corrected residuals.

TrainingUpdatesU criterionV criterionMax. absolute residual
U1,000−8.76×1011-8.76\times10^{11}1.20×10131.20\times10^{13}1.14×1091.14\times10^9
U3,000−1.82×1012-1.82\times10^{12}2.48×10132.48\times10^{13}1.64×1091.64\times10^9
V1,0000.04480.435146
V3,000-0.003850.11556.1

Under U-training, U decreases while the squared empirical kernel moment (V) and maximum corrected residual increase. V-training reduces both. This illustrates a source of the observed U instability.

3. Finite-sample bias and asymptotic consistency. V has finite-sample diagonal bias. For a fixed bridge with oracle nuisances, it is of order O(N−1)O(N^{-1}) under suitable moment conditions and vanishes asymptotically. Fitted nuisances introduce additional estimation errors, explicitly controlled through cross-fitting and the conditions in Appendix C. Eq. (40) defines normalized U and V losses, and the corresponding results establish recovery of the full-data population risk under their respective assumptions, including additional conditions for U where required. Finite-sample unbiasedness is therefore unnecessary for V consistency.

4. Comparison with prior work. Kompa et al. (2022, Fig. 2) report lower causal MSE for U in the standard fully observed Demand experiment, but greater robustness of V to increasing proxy noise (Sec. 6.2, Fig. S10). The regularized V criterion also follows PMMR (Mastouri et al., 2021). These results suggest design-dependent finite-sample bias-variance trade-offs, not universal superiority under MAR proxy missingness. We choose V because its unpenalized objective directly measures the corrected empirical moment norm.

Revisions. Section 3.1 discusses the V-statistic choice. Appendix E.1.6 presents the U/V comparisons, optimization diagnostic, and analyses (Tables 4-5).

AuthorsOpenReview ↗

Response to Critical Weakness 3/Question 2

Critical Weakness 3 & Question 1.2: Assumption D.6

Following the reviewer's suggestion, we developed an alternative DFPV causal-mean consistency proof using a dual treatment bridge, without requiring Assumptions D.6 or D.7(ii). We retain the original inverse-stability approach as an alternative set of sufficient conditions.

1. Strength of the original conditions.

Assumption D.6 bounds the distance to the valid-bridge set by the projected conditional-moment error:

d2(h,H0)≤τn∥E{Y−h(W,A,X)∣L}∥L2(PL).d_2(h,\mathcal H_0)\leq \tau_n\|\mathbb{E}\{Y-h(W,A,X)\mid L\}\|_{L_2(P_L)}.

Although this accommodates nonunique bridges, it remains strong. In ill-posed settings, τn\tau_n may grow with the neural class, and neither completeness nor neural parameterization guarantees this stability. Assumption D.7(ii) further transfers bridge-set convergence to the causal mean at fixed treatment aa, introducing a potentially growing factor Ca,nC_{a,n}. Together, they may require accurate recovery of bridge directions that are difficult to estimate but unnecessary for the target causal mean.

2. Direct recovery through a local dual bridge.

Building on Deaner (2023) and Bozkurt et al. (2026), we establish causal-mean consistency directly from projected bridge convergence, without requiring strong recovery of the outcome bridge.

For a fixed continuous treatment value, integrated L2(PA)L_2(P_A) convergence does not imply pointwise consistency. We therefore introduce local conditions requiring a square-integrable dual treatment bridge, positive treatment mass near the target, and Hölder regularity of the fitted causal-mean error. These transfer projected bridge convergence to consistency at the specified treatment value.

The proof also accounts for Stage 1 feature recovery and held-out averaging errors in the actual ProxMAR-DFPV estimator. The dual bridge is a population representer used only in the proof; it need not be estimated, and the estimator remains unchanged.

3. Quantitative comparison for neural classes.

Suppose the projected bridge error is Op(rn)O_p(r_n). For the same learning-error bound, the original inverse-stability route requires

max⁡{1,Ca,n}τnrn→0,\max\{1,C_{a,n}\}\tau_n r_n\to0,

whereas the local dual-bridge route requires

mnrn2s→0,m_n r_n^{2s}\to0,

where mnm_n controls the local ss-Hölder regularity factor of the fitted mean error.

For DFPV, this factor can be bounded using the coefficient norm, treatment-feature Hölder constants, and proxy/covariate-feature moments. Treatment-network regularity can also be controlled through layer spectral norms. If these factors are uniformly bounded, any vanishing projected error suffices for the new target-mean transfer step, without needing to offset a growing τn\tau_n.

4. Relationship between the guarantees.

The dual-bridge route avoids strong outcome-bridge recovery but requires dual-bridge existence and integrability, local treatment mass, and mean-curve regularity. Neither route universally dominates the other. The inverse-stability route additionally establishes strong bridge-set consistency, whereas the dual route directly targets the causal mean.

Revisions. Theorem 2 and Appendix D.4 establish causal-mean consistency under either set of sufficient conditions. Appendix D.3.4 specifies the local dual-bridge assumptions, their relation to neural feature maps, and the quantitative comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.