{"id":"10a62e00-d052-47c4-801c-61645c93ec47","arxiv_id":"1908.01920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An importance-weighting estimator for offline policy evaluation that provably achieves root-n consistency using proxies for latent confounders, without fitting an outcome model.","lead":"This paper develops a weighting method for evaluating new policies from logged data when hidden confounders exist but proxy variables are observed. It proves consistency under an adversarial balance objective, avoiding the need to fit an outcome regression.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5 and Algorithm 1 state a linear coefficient a_i that swaps indices relative to the proof in A.7; the printed QP does not minimize the adversarial objective, so the algorithm as written lacks the claimed consistency guarantee.","rationale":"The reader's weakest assumption (oracle for φ) is real but explicitly scoped by the paper ('we just assume we have some oracle'). The more damaging flaw is internal: the QP stated in Theorem 5 and Algorithm 1 does not match the derivation in its own proof. Appendix A.7 gives the linear term as -2/n^2 Σ_j W_j (Σ_i Q_ij π_{T_j}(X_i)), which is the coefficient of W_j with π evaluated at X_i using treatment T_j; after renaming, the coefficient of W_i is 2/n^2 Σ_j Q_ij π_{T_i}(X_j). The theorem statement instead has a_i = 2/n^2 Σ_j Q_ij π_{T_j}(X_i), swapping both the covariate and the treatment index. This means Algorithm 1 as printed minimizes a different quadratic form, so the returned weights are not the adversarial balancer of Eq. (7). The proof is correct, so the error is a typo, but it directly affects the paper's central algorithmic claim. Our concrete test compares the two formulae on the paper's own simulation; if the printed formula fails to balance, the paper must be corrected. We keep the reader's CONDITIONAL verdict because the method is salvageable by fixing the index, but the condition should include this correction. The oracle assumption remains a separate limitation, but it is not an internal inconsistency.","tokens_in":27236,"tokens_out":28902,"duration_ms":309011,"concrete_test":"Recompute the linear coefficient from the expansion in Appendix A.7 and compare it with the printed a_i in Theorem 5. Then, on the Section 5.1 synthetic scenario (Z~N(0,1), X 10-dimensional, m=2), generate a dataset of size n=2000, draw B=50 posterior samples per unit using the known φ, compute Q, and solve the QP twice: once with the printed a_i (π_{T_j}(X_i)) and once with the corrected a_i (π_{T_i}(X_j)). For both weight vectors, estimate the true sup_{μ∈F_K} J(W, μ) via a high-accuracy Monte Carlo over μ in the kernel ball and compare with the QP objective; then estimate the policy value bias over 64 replications of Y|Z,T. If the printed-a weights do not minimize the adversarial objective and their bias does not converge to zero as n grows, while the corrected-a weights reproduce the reported OptZ results, the algorithm as stated is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5 and Algorithm 1 define the linear term of the QP as a_i = 2/n^2 Σ_j Q_ij π_{T_j}(X_i). But expanding the adversarial objective in Appendix A.7 gives the coefficient of W_j as 2/n^2 Σ_i Q_ij π_{T_j}(X_i), which by symmetry is 2/n^2 Σ_j Q_ij π_{T_i}(X_j) for the coefficient of W_i. The printed a_i swaps the indices: it uses π at the covariate of unit i with the treatment of unit j, rather than π at the covariate of unit j with the treatment of unit i. Consequently the QP minimized by Algorithm 1 is not the supremum over F_K of J(W, μ) claimed in Theorem 5; the weights it returns do not solve Eq. (7), and Theorem 3's consistency guarantee does not apply to the algorithm as written. The correct formula is in the proof (A.7), so the error is fixable, but as printed the abstract's claim that Algorithm 1 'provably guarantees consistency' is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies off-policy evaluation of contextual bandit policies from logged data in the presence of latent confounders. The authors assume that the confounders Z render potential outcomes independent of (X,T), that X is a proxy for Z, and that a correct model of the posterior density φ(z;x,t) is available as an oracle. They first prove that no single importance weight can be unbiased for all outcome models (Theorem 1), then propose to select weights by minimizing the worst-case conditional mean squared error over a class F of candidate outcome regressions. Under regularity and complexity assumptions on F and assuming the true regression lies in F, they prove the resulting estimator is root-n consistent (Theorems 2 and 3). For RKHS classes they show the adversarial objective reduces to a quadratic program (Theorem 5) and provide Algorithm 1, estimating the Gram matrix by posterior sampling. Synthetic experiments compare the method with IPS, ignorability-based balancing, and direct regression methods.","tokens_in":27463,"tokens_out":21784,"duration_ms":200917,"significance":"This is a relevant and potentially useful contribution. The impossibility result (Theorem 1) clarifies why proxy-based balancing differs from standard IPS, and the adversarial-balance formulation, together with the consistency theorem (Theorem 3), offers a principled way to avoid fitting outcome regressions when confounders are latent. The appendix contains detailed proofs, and the experiments, while on a single synthetic scenario, show the expected qualitative behavior. However, the printed algorithm does not implement the theorem's objective: the Q-matrix estimator in Algorithm 1 uses the wrong posterior samples, and the linear coefficient in Theorem 5 and Algorithm 1 has a load-bearing index swap. In addition, a bound in the proof of Theorem 2 is not valid as stated. These issues are fixable but are central to the claim that Algorithm 1 'provably guarantees consistency,' so the paper needs a major revision.","major_comments":[{"comment":"The printed formula computes Q_ij = (1/B^2) ∑_{b=1}^B ∑_{c=1}^B K(Z_i^b, Z_i^c), using posterior samples from unit i in both arguments. This estimates E[K(Z_i, Z_i')], not the off-diagonal kernel expectation Q_ij = E[K(Z_i, Z_j')] defined in Theorem 5, where Z_j' is drawn from the posterior of unit j. The correct estimator, matching the proof in Appendix A.7, is (1/B^2) ∑_{b,c} K(Z_i^b, Z_j^c). As printed, Algorithm 1 does not minimize the adversarial objective, so the consistency guarantee of Theorem 3 does not apply to the weights it produces.","section":"Section 4.2, Algorithm 1, line 4"},{"comment":"The linear term is printed as a_i = (2/n^2) ∑_{j=1}^n Q_ij π_{T_j}(X_i). Expanding the objective in Appendix A.7 gives the coefficient of W_j as (2/n^2) ∑_{i=1}^n Q_ij π_{T_j}(X_i); using symmetry of Q, this is equivalent to a_i = (2/n^2) ∑_{j=1}^n Q_ij π_{T_i}(X_j). The printed version evaluates the policy of unit j at the covariate of unit i, whereas the correct term evaluates the policy of unit i at the covariate of unit j. Consequently the QP solved by Algorithm 1 is not the supremum over F_K of J(W, μ), and the printed claim that Algorithm 1 provably guarantees consistency is unsupported.","section":"Section 4.2, Theorem 5 and Algorithm 1, line 5"},{"comment":"The step bounding V[∑_t f_it μ_t(Z_i) | X,T] by max_t f_it^2 V[μ_t(Z_i) | X,T] does not follow from the lemma stated immediately above, which requires coefficients that sum to one; here ∑_t f_it = W_i - 1. The claimed inequality is false in general (for example, two negatively correlated variables with equal-magnitude coefficients give variance exceeding max_t f_it^2). A correct argument introduces sums of squares of f_it or (W_i - 1)^2 terms, changing the constants in J. The O(1/n) conclusion is likely still recoverable, but the proof as written is not valid.","section":"Appendix A.2, proof of Theorem 2"}],"minor_comments":[{"comment":"The word 'independant' should be 'independent'.","section":"Section 2.1, Assumption 1"},{"comment":"The algorithm omits the global 1/n^2 factors in G and a_i that appear in Theorem 5; this is equivalent to multiplying the whole objective by n^2 and leaves the minimizer unchanged, but the convention should be stated so that the hyperparameter Γ is interpreted consistently.","section":"Section 4.2, Algorithm 1"},{"comment":"In the derivation of sup_{μ_t ∈ F^K_t}, the display containing ⟨T_K s, T_K s⟩_K^2 / ||T_K s||_K should instead be ||T_K s||_K^2; the following line is correct, so this appears to be a typographical slip.","section":"Appendix A.7"},{"comment":"The sentence 'Some recent work on this problem involves using techniques such as matrix completion (Kallus, 2018a)' cites Kallus (2018a), which is the balanced policy evaluation paper; the relevant matrix-factorization reference appears to be Kallus, Mao, and Udell (2018).","section":"Section 2.3"},{"comment":"The claim that Assumptions 4 to 10 are satisfied by 'families of neural networks' is stated without proof or reference; only the RKHS case is established in Section 4.","section":"Section 3.3"},{"comment":"The caption reads 'Convergence of RMSE for for policy evaluation'; the duplicated 'for' should be removed.","section":"Section 5.2, Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is promising. The mismatches between the printed Algorithm 1 and the proof in A.7 are likely local errors, but because the abstract makes a strong 'provably guarantees consistency' claim about the algorithm, the authors must correct the Q and a formulas and ensure the released code matches the corrected version. I also recommend asking them to repair the variance bound in the proof of Theorem 2; the conclusion appears salvageable, but the proof as written is not correct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: substantive theory, real contribution, but the algorithm as printed does not implement the theorem. In Theorem 5 and Algorithm 1, the linear coefficient a_i is defined with swapped indices, and step 4 of Algorithm 1 estimates Q_ij using samples only from unit i instead of from both i and j. Both are clearly typos—the correct forms are in the proof in A.7—but as printed, the QP being solved is not the adversarial objective analyzed in A.7, so Theorem 3's consistency guarantee does not attach to the algorithm as written. This is a fixable presentation error, not a flaw in the underlying theory.\n\nWhat is actually new: Theorem 1 is a clean impossibility result—no single weight function can be unbiased for all outcome models when only proxies are available. The proposed remedy, adversarial balancing over the latent posterior, genuinely extends Kallus (2018a) from the unconfounded setting to latent confounders, and the proof that this achieves root-n consistency under RKHS classes (Theorems 3–5, plus the empirical-process arguments in A.5) is a real technical achievement. The appendix is thorough, the main proofs check out, and the experiments are an honest proof-of-concept: they assume the posterior oracle and show the method beating ignorability-based baselines.\n\nSoft spots, in proportion: the index bugs are the load-bearing issue because the paper's central claim is that Algorithm 1 'provably guarantees consistency.' As printed, that claim is unsupported. The fix is straightforward, but the experiments should also say which version of the algorithm was actually run. The other soft spot is the oracle assumption for the latent posterior φ(z;x,t) and the requirement that μ be in F. The paper is upfront that it does not address identification of the latent model, so this is a stated boundary rather than a hidden flaw—but it does limit the practical reach.\n\nWho this is for: anyone working on offline policy evaluation, causal inference with proxies, or balancing weights. The theory deserves a serious referee; the bugs are exactly what referees are for. Send it to peer review, with instructions that the algorithm/theorem mismatch must be corrected.","headline":"The proxy-balancing theory is a real step forward, but the printed algorithm has index errors that disconnect it from the consistency theorem; fixable, but as written the abstract's claim is unsupported.","tokens_in":27964,"tokens_out":4160,"would_cite":true,"duration_ms":40443,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that latent confounding can be overcome in offline policy evaluation by weighting observations to adversarially balance a class of mean outcome functions that is assumed to contain the truth.","keywords":["policy evaluation","latent confounders","proxy variables","importance weighting","adversarial balance","off-policy evaluation","reproducing kernel Hilbert space","conditional mean square error"],"falsifier":"In the paper's own simulation, replace the true posterior density $\\phi(z;x,t)$ with a deliberately wrong one—for instance, omit treatment from the conditioning or shift the posterior mean—and check whether the bias of the optimal balancing weights fails to shrink as $n$ grows; that would confirm that the oracle assumption carries the consistency guarantee.","tokens_in":27029,"feed_emoji":"⚖️","tokens_out":11447,"duration_ms":105142,"temperature":0.7,"pith_summary":"Offline policy evaluation usually needs the no-unobserved-confounding assumption, which often fails in practice. This paper tries to establish that consistency is still possible when the analyst has proxies for latent confounders and an oracle for their conditional density given observables and treatment. It first proves that no single importance-weight function can be unbiased for all outcome models in this setting, then shows that an adversarially chosen weighting scheme can nonetheless drive the conditional mean square error to $O(1/n)$, yielding a root-$n$ consistent estimator of policy value. If the claim is right, data logged under one treatment rule can be re-weighted to evaluate new policies in medicine and other high-stakes settings without fitting outcome regressions.","feed_headline":"Latent confounders no longer block policy evaluation","feed_subtitle":"Adversarial balance over proxy variables yields root-n consistent estimates from logged data.","key_machinery":"The central object is the adversarial balance objective $J(W,\\mu) = \\left(\\frac{1}{n}\\sum_{i=1}^n \\sum_{t=1}^m f_{it}\\nu_t(X_i,T_i)\\right)^2 + \\frac{2\\sigma^2}{n^2}\\|W\\|_2^2$, with $f_{it}=W_i\\delta_{T_i t}-\\pi_t(X_i)$; Theorem 2 bounds the conditional MSE of the weighted estimator by $2J(W,\\mu)+O(1/n)$. The paper minimizes $\\sup_{\\mu\\in\\mathcal{F}} J(W,\\mu)$ rather than the unknown $J(W,\\mu)$. The proof machinery is a minimax interchange, a dual calculation showing the minimum-norm weights that zero the bias term have controlled norm under Assumptions 4 to 10, and Rademacher-complexity control of the uniform deviation $\\sup_{\\mu\\in\\mathcal{F}} |E_n[\\nu_T(X,T)^2]-E[\\nu_T(X,T)^2]|$. For implementation, when $\\mathcal{F}$ is an RKHS ball the supremum over $\\mu$ becomes a quadratic form in the matrix $Q_{ij} = E[K(Z_i, Z_j')]$ with $Z_i,Z_i'$ shadow draws from the posterior density $\\phi(z;x_i,t_i)$, turning the objective into a quadratic program.","core_discovery":"Under Assumptions 4 to 10, with the true mean outcome vector $\\mu$ in a function class $\\mathcal{F}$, the adversarially balanced estimator $\\hat{\\tau}^{\\pi}_{W^*}$ satisfies $\\hat{\\tau}^{\\pi}_{W^*} = \\tau^{\\pi} + O_p(1/\\sqrt{n})$ (Theorem 3). The objective $J(W,\\mu)$ upper-bounds the conditional MSE of the weighted estimator, and the key step is showing $\\inf_W \\sup_{\\mu\\in\\mathcal{F}} J(W,\\mu) = O(1/n)$ via a minimax exchange and uniform convergence over $\\mathcal{F}$. The paper also proves an impossibility result: with latent confounding, weights that are unbiased for every outcome model would have to depend on the unknown outcome regressions $\\nu_t(x,t')$, so no universal IPS-style weights exist. When $\\mathcal{F}$ is the unit ball of a bounded Mercer reproducing kernel Hilbert space, the supremum over $\\mathcal{F}$ becomes a closed quadratic form and the weights come from a quadratic program.","pith_inferences":["The oracle assumption is doing real work: if $\\phi(z;x,t)$ is estimated from the same logged data, Theorem 3 does not by itself cover estimation error, so the practical guarantee depends on identifiability of the latent model.","A natural extension is to feed the same adversarial objective with posterior samples from more flexible fitted models and quantify how error in the posterior propagates into $Q$ and into the final estimate.","The impossibility result suggests that methods using only observed covariates $X$ cannot achieve universal unbiasedness under this causal graph; some proxy-model assumption is unavoidable.","The same minimax objective could be instantiated with other function classes, such as neural networks, by replacing the closed-form RKHS supremum with an inner-loop adversarial fit."],"forward_implications":["If the class $\\mathcal{F}$ contains the true outcome regression, weighted policy evaluation is consistent even when unconfoundedness given $X$ fails.","No fixed weighting recipe can be unbiased for all outcome models once confounders are latent; unbiasedness would require knowing the outcome regressions, so adversarial balance is the workable alternative.","With a bounded Mercer kernel such as the Gaussian kernel, the adversarial objective becomes a tractable quadratic program, making the method usable on real datasets.","In the paper's simulations, benchmarks that assume ignorability given $X$ show nonzero asymptotic bias, while bias of the balancing weights shrinks with sample size.","The same weights plug into doubly robust estimators, and consistency of those estimators follows whenever the weighting side is consistent."],"supporting_citations":[{"why":"Introduces adversarial balancing for policy evaluation in the unconfounded case; the proposed method ports this objective to latent confounders and uses it as the benchmark.","marker":"Kallus (2018a)"},{"why":"Defines the doubly robust estimator the proposed weights can be plugged into.","marker":"Dudík et al. (2011)"},{"why":"Supplies the lemma that converts $O(1/n)$ conditional mean square error into root-$n$ consistency.","marker":"Kallus (2016)"},{"why":"Provides the Rademacher comparison inequality used in the uniform-convergence proof.","marker":"Ledoux & Talagrand (2013)"},{"why":"Establishes finite covering numbers for Mercer kernel classes, used to prove the RKHS class is compact.","marker":"Zhou (2002)"},{"why":"Gives the vanishing Rademacher complexity of RKHS unit balls, yielding Assumption 10 for the kernel class.","marker":"Mendelson (2003)"}],"fun_headline_variants":["Adversarial balance: no more latent confounder bias","Proxies + adversarial balance = consistent policy eval","Latent confounders? Balance them out","No universal weights, but adversarially consistent","Policy eval without outcome models, despite confounders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee collapses if the oracle for the conditional density $\\phi(z;x,t)$ of the latent confounders given observed covariates and treatment is misspecified, because the balancing objective then no longer controls the true mean outcome functions and bias need not vanish.","fun_headline_variants_meta":{"raw":{"variants":["Adversarial balance: no more latent confounder bias","Proxies + adversarial balance = consistent policy eval","Latent confounders? Balance them out","No universal weights, but adversarially consistent","Policy eval without outcome models, despite confounders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1670,"prompt_tokens":867,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":742}},"tokens_in":483,"tokens_out":803,"duration_ms":8264,"temperature":1.0,"reasoning_tokens":742,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:00:22.939167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the paper's own simulation, replace the true posterior density $\\phi(z;x,t)$ with a deliberately wrong one—for instance, omit treatment from the conditioning or shift the posterior mean—and check whether the bias of the optimal balancing weights fails to shrink as $n$ grows; that would confirm that the oracle assumption carries the consistency guarantee.","supporting_citations":[{"cited_title":"The covering number in learning theory","cited_arxiv_id":null,"evidence_quote":"Establishes finite covering numbers for Mercer kernel classes, used to prove the RKHS class is compact."},{"cited_title":"On the performance of kernel classes","cited_arxiv_id":null,"evidence_quote":"Gives the vanishing Rademacher complexity of RKHS unit balls, yielding Assumption 10 for the kernel class."}],"review_version":1}