{"id":"fc60a1df-bd92-43aa-be15-a9241bc9ee4c","arxiv_id":"2509.00196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A reweighted quasi-likelihood plus PCA projection estimator removes hidden-variable bias in multivariate GLMs, with convergence rates and Berry-Esseen bounds.","lead":"This paper introduces G-HIVE, a way to estimate effects in multivariate response models when hidden variables bias the results, even for binary or count outcomes. The method reweights the data, subtracts the hidden-variable direction, and builds confidence intervals with a validated approximation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Berry-Esseen bound's vanishing requires unstated ||F^*v||_2=O(1), which the paper's own simulation scaling violates; the delta_2 term then diverges, so the inference guarantee is not established.","rationale":"The paper's main contribution is a new reweighted estimating-equation construction plus PCA projection to remove first-order bias, and the bias-rate results (Theorem 1) are coherent and well-supported by the proof sketches. The reader's weakest assumption, Assumption 1 (PB Theta = 0), is indeed important but it is an identifiability condition explicitly acknowledged by the authors; moreover, their inference is framed for the surrogate P_B^perp F^*, so even if Assumption 1 fails, the inferential statement about the surrogate remains valid. A more technical and, in my view, more load-bearing concern is the Berry-Esseen bound in Theorem 5. The simplified display (32) and the subsequent scaling conditions (rp<1/3, rM>1/2+rp) are claimed to show the BE bound goes to zero, but this relies on the unstated assumption ||F^*v||_2=O(1). The paper's own simulation DGP normalizes each row of Theta to unit norm, which gives ||F^*v||_2 ~ sqrt(M) for a fixed unit vector v. Substituting this into delta_2 yields a term sqrt(n/p) that diverges under p=o(n). Thus the theoretical support for the approximate confidence intervals does not cover the regime that the paper claims to analyze. This is not a claim that the method is wrong; the theorems are correct as upper bounds, but the inference guarantee is substantially weaker than presented. A concrete Monte Carlo check can settle whether the concern lands. The verdict should remain CONDITIONAL: the paper is promising and the estimation results are solid, but the inferential guarantee needs either a sharper bound or an explicit, satisfied assumption on ||F^*v||_2.","tokens_in":48004,"tokens_out":16663,"duration_ms":188958,"concrete_test":"Reproduce the Section 5.1 DGP with growing dimensions, e.g., n=10^4, p=n^{0.2}, M=n^{0.8}; generate A, B, Theta as described (rows normalized to unit norm and projected to satisfy Assumption 1), compute F^* approximately by solving (5) on a very large Monte Carlo sample (as in Section 5.3), and evaluate ||F^* e_1||_2. If it scales as sqrt(M) ~ n^{0.4}, then the first term in delta_2 is R_n sqrt(n/p) ~ n^{0.4}, which diverges, showing that the stated BE bound is not o(1) and the asymptotic coverage guarantee is not delivered. Alternatively, check analytically whether any assumption implies ||F^*v||_2=O(1); if none does, the simplified BE condition is vacuous for this design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference claim rests on Theorem 5's Berry-Esseen bound. In Section 4 (after Theorem 5), the authors simplify the bound to (32) by 'further assum[ing] Rn=O(1), K=O(1), and ||F^*v||_2=O(1)'. But ||F^*v||_2=O(1) is not part of Assumptions 1-4 and is not implied by them. Under the paper's own simulation DGP (Section 5.1), each row of Theta is normalized to have L2 norm 1, so for a typical unit vector v (e.g., v=e_j), ||F^*v||_2 ~ ||Theta v||_2 ~ sqrt(M) because F^*=Theta+O(1/sqrt(p)) row-wise. Plugging ||F^*v||_2=sqrt(M) into delta_2 = R_n ||F^*v||_2( sqrt(n/(pM)) + p sqrt(log(p∨M)/M) + p sqrt(n)/M ) gives a leading term R_n sqrt(n/p), which diverges whenever p=o(n), the regime the paper considers. Thus delta_2 -> infinity and the BE upper bound does not go to zero under the scaling (e.g., M~n^{rM}, p~n^{rp}, rp<1/3, rM>1/2+rp) that the authors give for the simplified bound. Consequently, the claim that the Berry-Esseen bound 'justifies the validity of the proposed approximate inference approach' (Abstract, Section 3.2, Discussion) is not actually supported for growing M unless the extra condition ||F^*v||_2=O(1) holds, which is not stated in the theorems and is violated by the paper's own simulation design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies multivariate-response generalized linear models with hidden variables under a factor-model assumption relating X to latent Z. It defines a pseudo-true parameter F* through inverse-variance reweighted estimating equations and proposes an estimator \\hat Θ = \\hat P_B^\\perp \\hat F, where \\hat F is a modified quasi-likelihood estimator and \\hat P_B^\\perp is obtained from PCA on reweighted residuals. The authors prove that the first-order approximation bias is O(1/√p) and the projected second-order bias is O(1/p), give convergence rates for \\hat F, \\hat P_B^\\perp, and \\hat Θ, and provide a Berry–Esseen bound for linear functionals of \\hat Θ centered at P_B^\\perp F*. Simulations and two real-data analyses illustrate the method against the naive MLE.","tokens_in":48489,"tokens_out":15216,"duration_ms":186304,"significance":"If the technical gaps noted below are addressed, this is a meaningful contribution. Extending surrogate-variable analysis and spectral deconfounding from linear regression to non-linear multivariate GLMs is non-trivial, and the proposed reweighted estimating equations are a sensible device to restore the orthogonality that fails for the ordinary misspecified GLM score. The explicit bias rates O(1/√p) and O(1/p) are interesting, and the paper is unusually candid about the fact that the inference target is the surrogate P_B^\\perp F*, not Θ itself. The proofs are detailed and the method is transparently algorithmic. The main reservations are that the stated vanishing Berry–Esseen bound requires an extra normalization on F* that is not part of the assumptions and is violated in the paper's own simulation regime, and that the sample-splitting argument in the proof of Theorem 5 is not fully matched to the final averaged estimator.","major_comments":[{"comment":"The simplified Berry–Esseen bound (32) is obtained by 'further assum[ing] Rn=O(1), K=O(1), and ||F^*v||_2=O(1)' (Section 4). The last condition is not part of Assumptions 1–4 and is not stated in Theorem 5. Under the simulation DGP of Section 5.1 (rows of Θ normalized to unit L2 norm) and the scaling M≍n^{rM}, p≍n^{rp} with rp<1/3, rM>1/2+rp, ||F^*v||_2 is of order √(M/p) for a generic unit v (e.g., v=e_j). Substituting this into δ2 gives a leading term Rn√n/p, which diverges because p=o(√n) in that regime. Hence the claim that the Berry–Esseen bound goes to zero and 'justifies the validity of the proposed approximate inference approach' (Abstract, §3.2, Discussion) is not established for growing M. The paper must either add the norm condition to the theorem statements and propagate it through the claims, or temper the asymptotic validity claim.","section":"§4, Theorem 5 / Eq. (32)"},{"comment":"The proof repeatedly says 'due to sample splitting, for simplicity we can equivalently assume that \\hat F_m is independent of Y_i and X_i'. But Algorithm 1 averages the two folds to produce \\hat F and \\hat Σ, and \\hat G_m in (20) is computed on all n observations with the averaged \\hat F_m. The averaged estimator is not independent of either fold, and the held-out independence used for the concentration steps is not satisfied by the final inferential object. The manuscript should spell out a cross-fitting scheme that supports the expansion in (71) and the variance estimator (31), or provide a proof under the actual dependence; otherwise the Gaussian approximation guarantee in Theorem 5 is not fully established.","section":"Appendix B.5, proof of Theorem 5 / Algorithm 1"},{"comment":"The confidence interval is for u^T P_B^\\perp F^*v, not for u^T Θv. The paper is transparent in Section 3.2 and Table 1, but the abstract and Section 7 phrase the result as general 'uncertainty quantification' and 'validity of the proposed approximate inference approach' without the surrogate caveat. Since the gap between P_B^\\perp F^* and Θ is only O(1/p) in Frobenius norm, for finite p it can exceed the parametric width; the favorable coverage for Θ in Table 1 is a finite-p simulation result rather than a theoretical guarantee. Please state plainly in the abstract and main text that the CI covers the surrogate P_B^\\perp F^*, and that no frequentist coverage claim is made for Θ.","section":"§3.2, §5.5, and Abstract"}],"minor_comments":[{"comment":"The displayed definitions say 'F^* := [F_1^{*T},...,F_M^{*T}]^T ∈ R^{M×K}' and similarly for \\hat F; the correct dimension is R^{M×p}, since each F_m is 1×p. Please fix this typo.","section":"§3.1, after (13)"},{"comment":"The y-axis label 'Scaled Squared Frobenius Error' is ambiguous. The text defines ||\\hat Θ − Θ||_F^2 / √(pM), but it would be helpful to state why this normalization is used and to show Monte Carlo error bars, especially since r ranges from 20 to 500.","section":"Figures 1 and 2"},{"comment":"The 'true' F^* is obtained by maximizing the modified quasi-likelihood with n=2×10^5. This is an approximation to the population F^*, not the exact estimand; please state this clearly and report Monte Carlo variability.","section":"§5.3"},{"comment":"Coverage probabilities are averaged over r=100 repetitions. At the 95% level the Monte Carlo standard error is about 0.04, which is not negligible. Report standard errors or use more repetitions, and separately report coverage for the surrogate P_B^\\perp F^* and for Θ.","section":"§5.5 / Table 1"},{"comment":"No code or data availability statement is provided. Given the many implementation choices (initial values for the non-concave quasi-likelihood, data-driven K estimation, eigenvalue ratio cutoff), releasing code would substantially strengthen the paper.","section":"Reproducibility"},{"comment":"The rate for \\hat s_n^2/n contains a term of the form 'p √(K log(p∨M)/n) log(M∨n)' inside brackets; the parentheses are difficult to parse. Please clean up the display and double-check that all logarithm powers are consistent with the proof.","section":"§4, Eq. (31)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of stat.ME and addresses a timely problem. The core estimation idea is sound and the authors are honest about the surrogate inference target. My main concern for the editor is that the inference claim in the abstract is stronger than what Theorem 5 actually delivers unless an additional normalization on F* is imposed, and that normalization is in tension with the M≫p regime the paper emphasizes. The sample-splitting gap in the proof of Theorem 5 is also worth asking the authors to close. Once these are repaired, the paper could be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper extends Bing et al. from linear to multivariate GLMs with hidden variables. The reweighted estimating equation plus projection is genuinely new, and the bias expansions in Theorems 1–4 are nontrivial and seem correct. The estimation side is solid: the approximation bias rates (1/sqrt(p) and 1/p) and the stochastic error rates are coherent, and the PCA step is well motivated. I would struggle to call the inference guarantee established, though. Theorem 5 gives a Berry–Esseen bound containing a term delta_2 that depends on ||F^*v||_2. Then, in the simplification after the theorem, the authors say they 'further assume' Rn=O(1), K=O(1), and ||F^*v||_2=O(1). That extra condition is not part of the theorem's assumptions and is not implied by their scaling. Their own simulation DGP normalizes each row of Theta to unit norm, so for typical unit v, ||F^*v||_2 is about sqrt(M/p), which grows under the stated rates M~n^{rM}, p~n^{rp} with rM>1/2+rp. Plugging that in, delta_2 does not go to zero; it diverges. So the claim that the Berry–Esseen bound 'justifies the validity of the proposed approximate inference approach' is not supported in the regime they analyze. The authors are transparent that they are adding a condition, but the abstract and the discussion around Theorem 5 give an unconditional impression. This is a repair, not a rewrite: the bound in (30) is correct as a statement about the Gaussian approximation error, but it needs a formal assumption like ||F^*v||_2=O(1) and a discussion of when it holds, or a sharper treatment that keeps the dependence on ||F^*v||_2 and shows it vanishes under the same scaling. Two smaller issues: no code or data, which limits easy verification, and Assumption 1 (PB Theta=0) is untestable but standard for this literature. On the whole this is a serious paper with a strong estimation story and an inference section that overstates its guarantee. It deserves a careful referee who will push on the Berry–Esseen condition. I'd bring it to a reading group on deconfounding or high-dimensional latent variable models, and I would likely cite it once the inference gap is fixed.","headline":"A substantial GLM extension of hidden-variable regression with a real gap between the claimed Berry–Esseen justification and the conditions under which it holds.","tokens_in":48869,"tokens_out":4647,"would_cite":true,"duration_ms":54144,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J12","62F12","62H25"],"pacs":[],"model":"deepseek-v4-flash","headline":"In multivariate GLMs with hidden confounders, an inverse-variance reweighting restores the linear-model structure, so that a spectral projection removes the first-order bias and yields Gaussian approximate inference for the projected parame","keywords":["generalized linear models","hidden variables","unmeasured confounders","multivariate response","modified quasi-likelihood","factor models","orthogonal projection debiasing","Berry-Esseen bound"],"falsifier":"Simulate a GLM with hidden confounders where Assumption 1 is violated by a controlled amount, forcing a nonzero component of Theta along a known direction in the column space of B, and check whether Theta-hat converges to Theta, to P_B-perp(Theta + BL), or to something else. A direct numerical check of the claimed rates is also decisive: compute F* and P_B-perp F* from a very large Monte Carlo sample and verify the O(1/sqrt(p)) and O(1/p) decay of the two per-response biases as p increases. A further check: in a setting with a measured confounder treated as hidden, compare G-HIVE interval cove","tokens_in":2200,"feed_emoji":"📊","tokens_out":2218,"duration_ms":109878,"temperature":0.7,"pith_summary":"Unobserved variables that influence both the covariates and the outcome bias ordinary regression, and the standard fix for linear models fails once the outcome model is nonlinear, because the bias no longer lies in the span of the hidden-factor loadings. This paper shows the failure has a remedy: fit each response with a quasi-likelihood whose score is divided by the model variance function b'', so that the leading bias of the fitted coefficient matrix F* again lies in the hidden-factor space and can be removed by the orthogonal projection P_B-perp. The remaining approximation bias per response is O(1/p), smaller the more covariates one has, and the projection matrix is recovered by PCA on the covariance of the reweighted residuals. Because the asymptotic bias still dominates the stochastic error at the sqrt(n) scale, exact inference on the original parameter Theta is replaced by second-order approximate inference on P_B-perp F*, with a Berry-Esseen bound controlling how close the Gaussian intervals are to correct. If right, this gives a practical pipeline for debiased estimation and uncertainty quantification in multivariate binary or count regression under hidden confounding.","feed_headline":"Reweight-then-project removes hidden-confounder bias in GLMs","feed_subtitle":"Dividing the score by the variance function lets a projection kill first-order bias, cutting error to O(1/p).","key_machinery":"The load-bearing device is the inverse-variance reweighted estimating equation E[(Y_m - b'(F*_m X))/b''(F*_m X) times X^T] = 0, equivalently a modified quasi-likelihood whose integrand divides by the variance function b''. This weight reshapes the leading misspecification bias from a term involving the diagonal matrix D, which does not lie in the column space of B and cannot be projected away, into B_m Sigma_Z A^T Sigma_X^-1, a row of the hidden loading matrix B, which the projection P_B-perp = I - B(B^T B)^-1 B^T annihilates. The companion mechanism is PCA: the covariance of the reweighted residuals takes the form diagonal plus B(.)B^T, so under the factor-model pervasiveness assumption its","core_discovery":"The paper's central claim is that the orthogonality property that makes hidden-variable correction work in linear regression, namely that the first-order misspecification bias lies in the column space of the hidden loading matrix B, can be restored in multivariate GLMs by replacing the ordinary score with an inverse-variance-weighted score. Define F* by E[(Y_m - b'(F*_m X))/b''(F*_m X) times X^T] = 0; then F*_m - Theta_m = B_m E(ZZ^T) A^T {E(XX^T)}^-1 + O(1/p), whose leading term is proportional to a row of B and is therefore killed by the projection P_B-perp. The paper proves the per-response norms ||F* - Theta||_F/sqrt(M) = O(1/sqrt(p)) and ||P_B-perp F* - Theta||_F/sqrt(M) = O(1/p), estim","pith_inferences":["Because Assumption 1 (P_B Theta = 0) is untestable from observed data, since B is never seen, a natural companion analysis is a sensitivity study: perturb the estimated projection toward directions of non-orthogonality and report how much Theta-hat and the coverage statements shift.","The reweight-then-project mechanism is a candidate template for other misspecified-model settings where bias has a structured column-space component; the paper's own open directions (general dependent noise, p > n via regularization) mark the natural next tests.","The Berry-Esseen regime conditions (r_p < 1/3 and r_M > (1 - r_p) or (1/2 + r_p)) imply a practical rule of thumb: the number of responses M must comfortably exceed the number of covariates p before the Gaussian approximation is trustworthy.","The deconfounding check demonstrated on the NHANES data suggests a general validation pattern: treat a measured variable as hidden, and compare G-HIVE's recovered coefficients with those from a model that observes it."],"forward_implications":["Because the residual bias after projection is O(1/p) per response, collecting more covariates actively improves the deconfounding: the approximation gap between the estimable target and Theta shrinks as p grows.","Confidence intervals built from Theta-hat's limiting Gaussian law cover the projected parameter P_B-perp F*, a second-order approximation of Theta, at the nominal level, with the Berry-Esseen bound quantifying the error; coverage for Theta itself is not guaranteed, and simulations show the naive MLE's intervals badly undercover.","When the number of responses M grows faster than the number of covariates p (roughly M asymp p^alpha with alpha > 2 in the paper's rate simplification), G-HIVE's convergence rate beats the naive MLE that ignores hidden variables, so more response variables are a blessing of dimensionality for estimating the projection.","The eigenvalue-ratio rule selects the number of hidden factors K without tuning parameters, making the pipeline fully data-driven, and simulation results show the data-driven choice performs close to the oracle versions that know K or the true projection.","The framework covers any GLM with bounded second derivative of the log-partition function (linear, logistic, Poisson), extending hidden-variable correction beyond the linear-regression settings previously solved."],"supporting_citations":[{"why":"The linear-regression predecessor: supplies the identifiability assumption P_B Theta = 0 and the PCA-on-residuals projection strategy that G-HIVE generalizes to GLMs.","marker":"(Bing et al., 2022)"},{"why":"Surrogate variable analysis, the origin of the hidden-variable and batch-effect modeling line that this paper extends beyond linear regression.","marker":"(Leek and Storey, 2007)"},{"why":"Large-dimensional factor model theory: provides the factor model (2) and the pervasiveness conditions in Assumption 4 used to recover the loading space from spectral decompositions.","marker":"(Bai, 2003)"},{"why":"Defines the pseudo-true parameter for misspecified likelihoods, which grounds the definition of F* and the bias expansion.","marker":"(White, 1982)"},{"why":"Introduces quasi-likelihood estimation, the object that the paper modifies into its inverse-variance-weighted version in equation (10).","marker":"(Wedderburn, 1974)"},{"why":"Supplies the Davis-Kahan variant used to convert the residual-covariance perturbation into the bound on the projection error.","marker":"(Yu et al., 2015)"},{"why":"Asymptotics for M-estimators with increasing dimension, the rate benchmark matched by the convergence rate of F-hat in Theorem 2.","marker":"(Portnoy, 1984)"},{"why":"Concentration bounds for sample covariance operators used in Lemma 4 to control the stochastic part of the estimated hidden-factor space.","marker":"(Koltchinskii and Lounici, 2017)"}],"fun_headline_variants":["Reweight, project, and conquer hidden bias in GLMs","Weighted scores make hidden-variable bias vanish in GLMs","Projection after variance-weighting kills GLM bias","Variance-weighted projection cuts GLM bias to O(1/p)","New estimator removes hidden confounder bias in GLMs"],"cache_read_input_tokens":50560,"weakest_assumption_plain":"The load-bearing premise is that the true regression coefficients are orthogonal to the column space of the hidden loadings (P_B Theta = 0); since B is never observed, this condition cannot be checked in the data, and if it fails the projection removes part of the genuine covariate-response association and the estimator targets a different quantity.","fun_headline_variants_meta":{"raw":{"variants":["Reweight, project, and conquer hidden bias in GLMs","Weighted scores make hidden-variable bias vanish in GLMs","Projection after variance-weighting kills GLM bias","Variance-weighted projection cuts GLM bias to O(1/p)","New estimator removes hidden confounder bias in GLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000869,"raw_usage":{"total_tokens":3658,"prompt_tokens":858,"completion_tokens":2800,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":2717}},"tokens_in":602,"tokens_out":2800,"duration_ms":21720,"temperature":1.0,"reasoning_tokens":2717,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:50:17.161756+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a GLM with hidden confounders where Assumption 1 is violated by a controlled amount, forcing a nonzero component of Theta along a known direction in the column space of B, and check whether Theta-hat converges to Theta, to P_B-perp(Theta + BL), or to something else. A direct numerical check of the claimed rates is also decisive: compute F* and P_B-perp F* from a very large Monte Carlo sample and verify the O(1/sqrt(p)) and O(1/p) decay of the two per-response biases as p increases. A further check: in a setting with a measured confounder treated as hidden, compare G-HIVE interval cove","supporting_citations":[],"review_version":1}