acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Population-Risk Reconstruction for Learning under Selective Labels

Abstract

Selective labels arise when outcomes are observed only for a nonrandom subset of the population, causing empirical risk minimization (ERM) to optimize prediction for the selected sample rather than the deployment population. We propose Adaptive Population-Risk Reconstruction (APR), a framework that exploits bridge observations: labeled cases that are unexpectedly selected yet resemble regions where outcomes are largely unobserved. APR combines selection surprise with affinity to the unlabeled population to construct continuous Target-relevance scores, and uses these scores to reconstruct separate Core and Target components of deployment-population risk. A global reliability term limits reconstruction when effective bridge information is weak, while an observation-specific gate restricts adaptation to inputs for which the reconstructed predictor is locally relevant. Bridge discovery is completed and frozen before outcome-model training. In progressive bridge-information simulations, APR recovers 94% of the AUC gap between ERM and a full-information oracle. Across six selective-label benchmark settings, APR achieves the highest mean Target-population AUC and the lowest mean log-loss among the compared methods. In a real-covariate semi-synthetic benchmark based on UCI Bank Marketing, APR also achieves the best mean Target- and full-population AUC, log-loss, and Brier score. Additional stress tests show robustness to bridge contamination, identification noise, moderate transportability violations, and moderate misspecification of the Target-population share, while performance degrades as bridge identification or outcome transportability approaches its limiting failure regime.

Primary area:
causal reasoning
Keywords:
selective labels, sample selection bias, distribution shift, risk reconstruction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YuQDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies learning under selective labels, where labels are missing non-randomly, causing standard empirical risk minimization (ERM) to underrepresent certain target populations. To address this issue, the authors propose Adaptive Population-Risk Reconstruction (APR). The method combines label-selection probabilities with similarity to the unlabeled data distribution to assess the relevance of labeled samples to the target population. These relevance scores are then used to weight labeled samples and estimate risks for the Core and Target populations separately. Finally, a global reliability weight and a local gating mechanism adaptively combine a standard predictor with a risk-reconstruction predictor. The paper also analyzes sufficient conditions and limitations of risk reconstruction. Experiments on synthetic and real-data-based semi-synthetic datasets show competitive performance against the compared baselines.

Critical strengths
  1. The paper addresses a practically relevant problem, focusing on the distribution mismatch between labeled training data and the target population caused by selective labeling.
  2. Leveraging a small number of labeled samples relevant to the target population for risk reconstruction provides a meaningful approach to mitigating selective-label bias.
  3. The theoretical analysis examines conditions under which risk reconstruction is valid and cases where identification is impossible, providing insights into the method's applicability.
Critical weaknesses
  1. Although the proposed Bridge identification mechanism has some novelty, the framework introduces several additional components, including complex sample weighting, global reliability estimation, and local gating. These components largely build on existing ideas in sample weighting and model ensembling, making the overall framework relatively complex while offering limited additional algorithmic novelty.

  2. Although Section 6 provides component ablations and a comparison with Fixed Bridge Reweighting, more direct simplified baselines are still missing, such as using the risk-reconstruction predictor alone without prediction fusion. This comparison would help distinguish the contribution of risk reconstruction from that of the additional fusion mechanism and better justify the complexity of the overall framework.

  3. The paper analyzes conditions for Bridge identification and consistency of the reconstructed risk, which has theoretical value. However, the consistency result relies on the estimated weights approximating reference weights that satisfy the risk-identification condition. It remains unclear under what conditions the proposed relevance scores and power-based weighting scheme satisfy this requirement. Therefore, the connection between the theoretical analysis and the practical method needs further clarification.

Questions for authors

Could the authors address the above weaknesses?

9 comments
AuthorsOpenReview ↗

Response to Weakness 1

We thank the reviewer for this helpful comment. We agree that several components of APR build on existing ideas in sample weighting and model ensembling. However, the main novelty of APR lies in using bridge observations to reconstruct population risk from selectively observed labels. Rather than simply correcting selection bias through sample weights, APR identifies labeled observations that may contain useful information about the under-observed Target population and uses them to reconstruct the Core and Target risk components. Importantly, bridge discovery relies on observed covariates and label-availability information, without requiring known Core/Target identities, additional outcome labels, or a fully specified outcome-dependent selection mechanism. This makes it practical in settings where collecting additional Target labels is difficult or costly. The remaining components serve distinct purposes. Global reliability controls how much we can trust the reconstructed information overall, while local gating determines where it should be used. These two mechanisms are complementary: one controls the overall strength of adaptation, and the other controls its use for individual predictions.

To test whether these components are useful, we added four ablation variants: APR-SimpleWeight, APR-NoGlobal, APR-NoLocal, and APR-NoFusion. The results show clear benefits from the full design in the tested Valid Bridge Stress setting. Compared with APR-SimpleWeight, full APR improves AUC by 0.02790.0279 on Test A and 0.01050.0105 on Test B. Removing global reliability also reduces performance. Likewise, full APR improves AUC over APR-NoLocal by 0.02780.0278 on Test A and 0.01150.0115 on Test B. The APR-NoFusion results reveal an interesting trade-off. On Test A, which contains only Target observations, removing fusion slightly improves AUC by 0.00530.0053. This is expected because the reconstructed predictor is especially suited to Target-only evaluation. However, full APR achieves better Log-Loss on Test A (+0.0089+0.0089 improvement). More importantly, on Test B, which represents the mixed deployment population, full APR improves both AUC (+0.0057+0.0057) and Log-Loss (+0.0065+0.0065). The results are consistent with the intended stabilizing role of fusion. Relative to NoFusion, full APR improves mixed-population AUC and Log-Loss and Target-only Log-Loss, while reducing Target-only AUC. This ablation establishes a performance trade-off in the tested setting, rather than directly isolating robustness to reconstruction error.

AuthorsOpenReview ↗

Response to Weakness 2

We thank the referee for raising this point on adding baselines. We consider four ablations to assess the contribution of APR's key components: APR-SimpleWeight replaces the population-risk reconstruction weights with simple relevance-based weights wi=1+5qiw_i=1+5q_i; APR-NoGlobal removes adaptive global reliability estimation by fixing the global mixing coefficient at γ=0.75\gamma=0.75; APR-NoLocal replaces the observation-specific local gate a(x)a(x) with a constant estimated from the development set; and APR-NoFusion removes the final prediction fusion and uses only the reconstruction predictor f^R\hat f_R. All other components are retained whenever applicable. The table below reports the performance differences between Full APR and its ablated variants under Valid Bridge Stress, where positive values favor Full APR.

The role of prediction fusion. An interesting but interpretable result is that APR-NoFusion achieves a slightly higher AUC on Test A (ΔAUC=−0.0053\Delta\mathrm{AUC}=-0.0053). This is consistent with the design of APR: Test A consists exclusively of Target observations, representing an extreme evaluation setting in which the reconstructed predictor is particularly well aligned with the test population. Consequently, incorporating the ordinary ERM predictor through fusion may slightly dilute the Target-specific ranking advantage of the reconstructed predictor. However, this does not imply that removing fusion improves overall predictive quality. In fact, Full APR achieves a lower Log-Loss on Test A (ΔLogLoss=+0.0089\Delta\mathrm{LogLoss}=+0.0089). This distinction arises because AUC measures ranking performance, whereas Log-Loss evaluates the quality of predicted probabilities and penalizes confident incorrect predictions. Fusion can therefore improve probability predictions even when it slightly reduces Target-specific ranking performance.

Importantly, Test A represents an extreme, Target-only evaluation setting, whereas Test B evaluates the mixed Core--Target population and is more representative of the population-level learning objective considered in this study. On Test B, Full APR outperforms APR-NoFusion in both AUC (Δ=+0.0057\Delta=+0.0057) and Log-Loss (Δ=+0.0065\Delta=+0.0065). These results highlight the practical motivation for prediction fusion: it serves as a protective mechanism that moderates reliance on the reconstructed predictor by retaining a contribution from ordinary ERM. This design aims to reduce vulnerability to imperfect bridge reconstruction and excessive Target-specific adaptation, rather than maximizing performance on a purely Target population. The ablation results support its practical value for population-level prediction, while also revealing a modest trade-off in Target-only AUC.

The contributions of the remaining components. Under the tested Valid Bridge Stress setting, these three ablations show measurable performance benefits from the full weight construction, global reliability adjustment, and local adaptation relative to their specified alternatives. Replacing the proposed reconstruction weights with simplified weights (APR-SimpleWeight) reduces performance across both test settings, with AUC differences of +0.0279+0.0279 on Test A and +0.0105+0.0105 on Test B. Removing global reliability adjustment (APR-NoGlobal) also consistently deteriorates performance, confirming the usefulness of accounting for the overall reliability of reconstructed Target evidence. Similarly, removing local adaptation (APR-NoLocal) produces substantial performance degradation, particularly in AUC (+0.0278+0.0278 on Test A and +0.0115+0.0115 on Test B), supporting the importance of observation-specific adaptation. All three ablations also exhibit higher Log-Loss than Full APR, with positive performance differences and 95% confidence intervals excluding zero across both test settings.

Overall, these results demonstrate that the proposed weight construction, global reliability adjustment, and local adaptation each provide measurable contributions to APR's predictive performance. Prediction fusion serves a complementary role, improving probability prediction and mixed-population performance while allowing a small reduction in Target-only ranking accuracy. Together, the ablations provide empirical justification for the distinct components of the full APR framework.

AblationAUC — Test AAUC — Test BLog-Loss — Test ALog-Loss — Test B
APR-SimpleWeight+0.0279 [+0.0206, +0.0351]+0.0105 [+0.0056, +0.0154]+0.0134 [+0.0089, +0.0179]+0.0056 [+0.0034, +0.0078]
APR-NoGlobal+0.0103 [+0.0088, +0.0119]+0.0033 [+0.0021, +0.0045]+0.0054 [+0.0041, +0.0068]+0.0021 [+0.0014, +0.0028]
APR-NoLocal+0.0278 [+0.0240, +0.0316]+0.0115 [+0.0086, +0.0144]+0.0133 [+0.0091, +0.0176]+0.0062 [+0.0042, +0.0082]
APR-NoFusion-0.0053 [-0.0060, -0.0046]+0.0057 [+0.0041, +0.0073]+0.0089 [+0.0048, +0.0130]+0.0065 [+0.0046, +0.0084]
AuthorsOpenReview ↗

Response to Weakness 3 (Part 1/3)

We thank the reviewer for this helpful comment. To clarify the connection between theory and implementation, we have developed additional theoretical results for APR's actual relevance-score transformation and power-based weighting scheme, without changing the algorithm. We summarize the main conditions, results, and proof arguments below. Detailed statements, technical assumptions, proofs, and illustrative constructions have been incorporated into the new revision of paper (Appendix B).

1. Conditions for valid relevance scores.

Let H=(X,G)H=(X,G), e(G)=Pr⁡(S=1∣G)e(G)=\Pr(S=1\mid G), and d(X)=Pr⁡(S=0∣X)d(X)=\Pr(S=0\mid X). APR constructs the population counterpart of its relevance score as

Q(H)=Sτ[αN((1−e(G))d(X))+(1−α)d~(X)].Q(H)=\mathcal S_\tau\left[\alpha\mathcal N\big((1-e(G))d(X)\big)+(1-\alpha)\widetilde d(X)\right].

We establish a score-validity lemma with two implications. First, if the selection-surprise and affinity signals (1−e(G),d(X))(1-e(G),d(X)) for Target observations multivariate first-order stochastically dominate those for Core observations conditional on S=1S=1, then

Pr⁡(Z=T∣S=1,Q(H)≥t)≥Pr⁡(Z=T∣S=1)\Pr(Z=T\mid S=1,Q(H)\ge t)\ge \Pr(Z=T\mid S=1)

for thresholds with positive tail probability.

Second, suppose a measurable Target-like region A\mathcal A satisfies

1−e(G)≥a+,d(X)≥d+,H∈A,1−e(G)≤a−,d(X)≤d−,H∈Ac,\begin{aligned}1-e(G)&\ge a_+,\quad d(X)\ge d_+,&& H\in\mathcal A,\\\\ 1-e(G)&\le a_-,\quad d(X)\le d_-,&& H\in\mathcal A^c,\end{aligned}

where a+>a−a_+>a_- and d+>d−d_+>d_-. If normalization and clipping preserve a strict separation between the transformed signals, then

Q(H)≥q+>q−≥Q(H′),H∈A,H′∈Ac.Q(H)\ge q_+>q_-\ge Q(H'),\qquad H\in\mathcal A,\\ H'\in\mathcal A^c.

Proof sketch. The product of the nonnegative bridge signals is coordinatewise nondecreasing, as are APR's robust normalization, prevalence adjustment, and sharpening. Stochastic dominance therefore preserves Target enrichment, while stronger signal separation yields a strict score gap when endpoint clipping is nondegenerate.

These conditions explain when APR's implemented score is informative. Importantly, Q(H)Q(H) is a relevance score, not a calibrated posterior probability of latent Target membership.

Reviewer_qqN2OpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a novel perspective on the problem of selective labels: rather than treating unobserved labels merely as "missing data" to be imputed or corrected for bias, it leverages "bridge observations", labeled samples that were incidentally selected and share characteristics with the unlabeled population, to reconstruct the overall risk for the target deployment population. The work is substantial in scope, and the writing is clearly structured.

Critical strengths
  1. It is a novel perspective to separate the selective label bias from the traditional paradigm of "missing mechanism modeling, reweighting, or imputation" and propose the complementary information source that "the selected samples themselves carry the label traces of the under-observed population".
  2. The experimental results in the paper demonstrate that the gains indeed stem from "risk reconstruction" rather than merely from bridging upsampling.
Critical weaknesses
  1. The paper incorporates an excessive number of hyperparameters, yet may fail to provide sufficient justification for its reliance on the hyperparameter tuning process.
Questions for authors
  1. How to choose the optimal hyperparameters, namely α\alpha, pTp_T, rTr_T, pCp_C, κ\kappa, ζ\zeta, and τ\tau?
  2. Unfortunately, I just forgot to bid on papers, and to be honest, I am not very familiar with this research field, so I do not have any specific questions for the authors.
1 comment

We thank the reviewer for raising this important concern regarding the number of hyperparameters and the justification for their selection. As already documented in Appendix K (Hyperparameter Choices) of the original manuscript, APR's structural hyperparameters have distinct and interpretable roles rather than serving as unrestricted tuning coefficients. Specifically, α\alpha balances direct bridge evidence and Target affinity; pTp_T, rTr_T, and pCp_C control the concentration of the reconstructed Target and Core masses; κ\kappa determines the growth of global reconstruction reliability; and ζ\zeta and τ\tau govern the sharpness of local adaptation and bridge relevance, respectively. The baseline configuration was selected once using an independent development sample of Nval=1,000N_{\mathrm{val}}=1{,}000, with small prespecified grids of four or five plausible values for each tunable parameter and development-set AUC as the selection criterion. Importantly, the selected configuration was subsequently held fixed across both Valid Bridge and Unified MNAR selection regimes, all three severity levels, and all evaluation seeds. Neither scenario-specific retuning nor test-set-based hyperparameter selection was performed. This procedure limits the risk that APR's reported advantages arise from excessive tuning to particular experimental conditions.

Furthermore, Appendix H (Hyperparameter Sensitivity Analysis, Figure 10 and Table 7) explicitly examines the extent to which APR depends on these choices. Using 30 Monte Carlo replications, we independently vary six principal hyperparameters around their baseline values while fixing pC=2p_C=2, and evaluate low-, medium-, and high-bridge-information regimes. We keep pC=2p_C=2 to maintain a simple and consistent Core-weighting rule, where observations with higher Target relevance receive progressively less Core weight. This allows the sensitivity analysis to focus on the six parameters that more directly control bridge identification, Target reconstruction, and reliability-aware adaptation.

The results show that APR is generally stable under moderate parameter perturbations, particularly when bridge information is sufficiently informative. The effects of α\alpha, rTr_T, and ζ\zeta are relatively small, while pTp_T exhibits an interpretable concentration trade-off and κ\kappa has a more noticeable effect when bridge information is scarce, reflecting the intended conservativeness of global reliability control. Across 30 Monte Carlo replications, the mean recovery rate varies by less than 3.6%3.6\% relative to baseline across all tested hyperparameter perturbations in the medium- and high-bridge-information regimes (Appendix H, Table 7). Sensitivity is greater when bridge information is scarce, particularly for κ\kappa, which controls how conservatively APR uses limited Target evidence. Overall, these results show that performance varies relatively little over the tested one-at-a-time perturbations around the baseline configuration. This supports limited local sensitivity in these settings, but does not establish robustness to arbitrary joint parameter choices or different data-generating processes.

Reviewer_2tEYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Adaptive Population-Risk Reconstruction (APR) for prediction when outcomes are selectively observed. APR combines estimated selection surprise and affinity to unlabeled covariates into a frozen relevance score, uses that score to construct separately normalized Core and Target training losses, and blends the resulting predictor with an ordinary ERM predictor through an effective-sample-size factor and a learned gate. The appendix gives a conditional enrichment result, an observational indistinguishability result, and a consistency argument assuming suitable reconstruction weights. Evaluation comprises progressive bridge-information simulations, two synthetic selection mechanisms at three severities, a Bank Marketing construction with imposed selective observation, and ablation and sensitivity experiments. The reported comparisons generally favor APR.

Critical strengths
  • Rich experiments, each error source (bridge, estimation, transport, mixture) having a dedicated experiment, and a negative-control identifiability experiment is nice addition. Empirical results show promising gains across ranking and probability scores.
Critical weaknesses
  • The enrichment proposition and the conditional ratio-estimator argument seem to be sound under their stated assumptions. But those assumptions are not established for APR’s implemented weights.
  • The canonical IPW and density-ratio baselines are missing from empirical comparison.
  • The novelty is overstated. I see this work as tempered propensity x density-ratio + gating, and these ingredients are pretty standard in this area (propensity weighting, classifier density ratio, mixture-of-experts gating).
Questions for authors
  • Can you provide complete experimental mechanisms? What are the synthetic outcome/selection equations, audit probabilities, bridge counts, and Bank Target-assignment and masking rules?
  • Do baselines also receive/use G?
7 comments
AuthorsOpenReview ↗

Response to Weakness 1— Part 1/3

We thank the reviewer for pointing out this important gap between our theoretical assumptions and APR's implemented weights. The original results establish conditional guarantees, but do not fully explain when the required conditions arise from the underlying data-generating process (DGP).

To address this, we provide a constructive DGP analysis illustrating when APR's observable signals are informative, when its power weights can approximate the desired distributional correction, and when this approximation may fail. We then connect these findings to new theoretical bounds for APR's unchanged implementation. We also update these in Appendix B of new revision of paper.

1. Constructive DGP conditions and failure modes.

Consider a latent Core/Target mixture with observed covariates H=(X,G)H=(X,G) and latent membership Z∈C,TZ\in\\{C,T\\}. Define ηT(H)=Pr⁡(Z=T∣H)\eta_T(H)=\Pr(Z=T\mid H).

As a simple constructive case, suppose the component-specific label-observation probabilities are constant within components:

Pr⁡(S=1∣H,Z=T)=sT,Pr⁡(S=1∣H,Z=C)=sC,\Pr(S=1\mid H,Z=T)=s_T,\qquad \Pr(S=1\mid H,Z=C)=s_C,

where 0<sT<sC<10<s_T<s_C<1. The resulting marginal non-selection probability is

Pr⁡(S=0∣H)=1−sC+(sC−sT)ηT(H).\Pr(S=0\mid H) = 1-s_C+(s_C-s_T)\eta_T(H).

Thus, higher latent Target prevalence implies higher observable non-selection probability. In the special case X=G=HX=G=H, APR's two auxiliary signals satisfy d(H)=1−e(H)=Pr⁡(S=0∣H)d(H)=1-e(H)=\Pr(S=0\mid H).

For two regions h+h_+ and h−h_- with ηT(h+)>ηT(h−)\eta_T(h_+)>\eta_T(h_-), the implemented relevance score satisfies Q(h+)>Q(h−)Q(h_+)>Q(h_-) whenever its monotone transformations preserve strict separation. Consequently, APR's Target mass UT=QpTdrTU_T=Q^{p_T}d^{r_T} assigns greater weight to h+h_+.

This illustrates a constructive connection from primitive selection conditions to APR's actual weighting direction, rather than assuming that the relevance score is informative.

However, correct ranking is not sufficient for accurate density-ratio reconstruction. The ideal Target weight is proportional to ηT(H)/Pr⁡(S=1∣H)\eta_T(H)/\Pr(S=1\mid H), whereas APR uses Q(H)pTd(H)rTQ(H)^{p_T}d(H)^{r_T}.

In the two-region case, the approximation is governed by the difference between their log-weight ratios:

εratio=∣log⁡UT(h+)UT(h−)−log⁡wT⋆(h+)wT⋆(h−)∣.\varepsilon_{\mathrm{ratio}} = \left| \log\frac{U_T(h_+)}{U_T(h_-)} - \log\frac{w_T^\star(h_+)}{w_T^\star(h_-)} \right|.

Under positive probabilities and weights, the corresponding normalized two-region distributions satisfy

TV⁡(νTU,PTH)≤εratio4.\operatorname{TV}(\nu_T^U,P_T^H) \le \frac{\varepsilon_{\mathrm{ratio}}}{4}.

Proof sketch. Please note that formal statements and proofs appear in the revised manuscript. In a two-region distribution, normalization converts weight ratios into probabilities through the logistic function, whose derivative is bounded by 1/41/4. The stated bound follows by the mean-value theorem.

In the two-region construction, a small log-weight-ratio discrepancy guarantees closeness of the induced region probabilities. This illustrates a calibration requirement for APR’s implemented weights, and reconstruction of the full component distribution additionally requires adequate support coverage and within-region agreement.

The same construction also reveals three failure modes. First, if sT=sCs_T=s_C, the bridge signals in this DGP become uninformative about Target prevalence. Second, even with correct relevance ranking, poorly calibrated powers may produce substantial distributional error. Third, if selected conditional outcomes differ from the intended component outcomes, accurate covariate weighting alone cannot recover the component risk.

These cases clarify both the usefulness and the limits of the assumptions examined below.

AuthorsOpenReview ↗

Response to Weakness 1— Part 2/3

2. Validity of the implemented relevance score.

Let e(G)=Pr⁡(S=1∣G)e(G)=\Pr(S=1\mid G) and d(X)=Pr⁡(S=0∣X)d(X)=\Pr(S=0\mid X). APR constructs

Q(H)=Sτ[αN((1−e(G))d(X))+(1−α)d~(X)].Q(H)=\mathcal S_\tau\left[ \alpha\mathcal N((1-e(G))d(X)) +(1-\alpha)\widetilde d(X) \right].

We establish a score-validity lemma showing that multivariate stochastic dominance of the selection-surprise and affinity signals preserves Target enrichment under APR's actual transformation.

Under the stronger condition that these signals uniformly separate a Target-like region A\mathcal A from its complement, and provided normalization and clipping do not eliminate this separation,

Q(H)≥q+>q−≥Q(H′),H∈A,H′∈Ac.Q(H)\ge q_+>q_-\ge Q(H'), \quad H\in\mathcal A,\\ H'\in\mathcal A^c.

Proof sketch. The product of the nonnegative bridge signals and APR's normalization, prevalence adjustment, and sharpening transformations are coordinatewise nondecreasing. They therefore preserve the stochastic-dominance enrichment implication and, under stronger nondegenerate signal separation, produce a strict score gap.

3. Structural approximation of implemented weights.

APR uses

UT(H)=Q(H)pTd(X)rT,UC(H)=(1−Q(H))pC.U_T(H)=Q(H)^{p_T}d(X)^{r_T}, \qquad U_C(H)=(1-Q(H))^{p_C}.

Let P1H=L(H∣S=1)P_1^H=\mathcal L(H\mid S=1) and PgH=L(H∣Z=g)P_g^H=\mathcal L(H\mid Z=g). The normalized weights induce νgU(dh)=Ug(h)P1H(dh)/mg\nu_g^U(dh)=U_g(h)P_1^H(dh)/m_g, where mg=E[Ug∣S=1]>0m_g=\mathbb E[U_g\mid S=1]>0. Define Δg=TV⁡(νgU,PgH)\Delta_g=\operatorname{TV}(\nu_g^U,P_g^H).

Under the score-separation condition, we obtain the structural approximation bound

ΔT≤KT1+KT+δT+ξT,ΔC≤KC1+KC+δC+ξC,\boxed{ \begin{aligned} \Delta_T&\le\frac{K_T}{1+K_T} +\delta_T+\xi_T,\\\\ \Delta_C&\le\frac{K_C}{1+K_C} +\delta_C+\xi_C, \end{aligned} }

where, for r=P1H(A)∈(0,1)r=P_1^H(\mathcal A)\in(0,1),

KT=1−rr(q−q+)pT(d−d+)rT,KC=r1−r(1−q+1−q−)pC.\begin{aligned} K_T&=\frac{1-r}{r} \left(\frac{q_-}{q_+}\right)^{p_T} \left(\frac{d_-}{d_+}\right)^{r_T},\\\\ K_C&=\frac{r}{1-r} \left(\frac{1-q_+}{1-q_-}\right)^{p_C}. \end{aligned}

Here d+>d−≥0d_+>d_-\ge0 are the affinity bounds on A\mathcal A and its complement, δg\delta_g measures missed component coverage, and ξg\xi_g measures within-region distributional mismatch.

Proof sketch. Score separation bounds the ratio of weighted mass outside and inside the component-relevant region by KgK_g. This yields cross-region leakage at most Kg/(1+Kg)K_g/(1+K_g); the total-variation triangle inequality then establishes the structural approximation bound above.

The bound directly connects APR's actual powers to structural approximation error. Larger powers reduce cross-region leakage under separation, but accurate reconstruction additionally requires adequate coverage and within-region balance.

Exact distributional reconstruction occurs if and only if Ug=cgwg⋆U_g=c_gw_g^\star almost surely, where cg>0c_g>0 and wg⋆=dPgH/dP1Hw_g^\star=dP_g^H/dP_1^H, assuming the required support conditions. This calibration requirement does not follow from enrichment alone.

We conduct an additional constructive simulation with same DGP to assess the practical relevance of the bound, using APR's relevance-score transformation and power-based weighting scheme. The table below reports the evaluated bound components and empirical Target reconstruction discrepancy across three severity levels. Across all three severity levels, the empirical discrepancy remains below the corresponding theoretical bound. In this construction, the cross-region leakage term is near-zero, indicating that APR's score separation and power-based weighting effectively suppress cross-region contamination in this particular setting. But the empirical Target distribution discrepancy remains 0.184-0.395 across severity levels. Thus, strong separation does not imply accurate reconstruction. The empirical discrepancies are below the evaluated bounds in all three settings.

Numerical verification of the structural approximation bound (20 seeds; mean values).

SeverityKT1+KT\frac{K_T}{1+K_T}δT\delta_TξT\xi_TTotal BoundEmpirical ΔT\Delta_T
Mild7.96×10−157.96\times10^{-15}0.08000.14220.22220.1843
Moderate8.07×10−148.07\times10^{-14}0.12000.18470.30470.2487
Severe4.27×10−124.27\times10^{-12}0.18000.29710.47710.3948
AuthorsOpenReview ↗

Response to Weakness 1— Part 3/3

4. Implications for risk reconstruction.

For a fixed predictor ff and bounded loss 0≤L(f)≤M0\le L(f)\le M, let Rg(f)R_g(f) be the true component risk and RgU(f)R_g^U(f) its population weighted counterpart. We establish

∣RgU(f)−Rg(f)∣≤MΔg+Tg(f),\boxed{ |R_g^U(f)-R_g(f)| \le M\Delta_g+\mathcal T_g(f), }

where

Tg(f)=EPgH∣E[L(f)∣H,S=1]−E[L(f)∣H,Z=g]∣\mathcal T_g(f)= \mathbb E_{P_g^H} \left| \mathbb E[L(f)\mid H,S=1] - \mathbb E[L(f)\mid H,Z=g] \right|

measures conditional outcome-transport error.

Proof sketch. Adding and subtracting the component expectation of the selected conditional loss separates structural weighting error from outcome-transport error. The former is controlled by total variation.

We also derive a finite-sample extension accounting for estimation errors in the relevance scores and their power transformations. It identifies a concentration--stability tradeoff: larger powers can reduce contamination but may amplify estimation errors and reduce effective component mass.

Scope and final revision.

The constructive analysis provides an explicit DGP class in which APR's bridge signals are informative, together with a quantitative characterization of weight calibration and identifiable failure modes. It does not establish that the calibration and transport conditions hold in every application.

Our theoretical guarantees are therefore conditional: score enrichment alone does not imply exact weighting, and accurate risk reconstruction additionally requires controlled structural mismatch and conditional outcome transportability. In particular, the analysis does not claim universal identification under unrestricted MNAR selection.

The formal propositions, technical assumptions, detailed proofs, and additional constructive analysis have been incorporated into the revised manuscript. These additions clarify when APR's implemented weights satisfy the theoretical reconstruction requirements, when they may fail, and the limits of the resulting guarantees.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.