{"id":"283d8f7a-d9dd-4a1c-b45b-e7f00e960e5a","arxiv_id":"2508.12688","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bayesian re-parameterization of the partially linear model recovers the causal effect from the reduced-form error covariance, avoiding regularization-induced confounding and matching double machine learning asymptotics.","lead":"This paper introduces a fully Bayesian method for estimating causal effects when many control variables are present, by modeling outcome and treatment together and reading the causal effect off their error correlation. It promises the frequentist guarantees of double machine learning with better finite-sample coverage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BvM theorem is proven only under known error variances, while the BDML procedure as implemented treats the full error covariance as unknown; the abstract's frequentist-validity and efficiency claims therefore go beyond what Proposition 7 establishes.","rationale":"The reader's weakest-assumption diagnosis is accurate: the BvM theorem is the load-bearing support for the abstract's inference and efficiency claims, and it is stated under known error variances plus an asserted-but-unverified adaptation of Walker (2025) to the empirical L2 norm. My reading of the full text confirms that Algorithm 1 and the simulation study use a full Bayesian treatment of the error covariance Σ, so Proposition 7's known-variance posterior is not the posterior whose frequentist properties the paper claims. Proposition 5 does show the BDML posterior mean has bias of order p²/n² and is √n-consistent for p=o(n^{3/4}), which is suggestive but not a BvM theorem; it does not establish asymptotic normality of the posterior or valid credible intervals. The proof sketch for Proposition 7 verifies some high-level conditions but does not derive the posterior of the ratio α=Σ12/Σ22, so the gap is real rather than merely cosmetic. This does not mean the method is wrong—the simulation evidence and bias calculations are promising—but the advertised theoretical guarantees are not proven for the actual procedure. The reader's conditional verdict is therefore appropriate, and my stress-test does not change it. I also note the abstract's promised empirical examples are absent, and no code/data are provided, which supports the conditional rather than unconditional acceptance.","tokens_in":30085,"tokens_out":7863,"duration_ms":90946,"concrete_test":"Run a focused simulation comparing the actual BDML posterior from Algorithm 1 (unknown Σ, inverse-Wishart/LKJ prior) with the known-variance posterior used in Proposition 7, under the DGP in (20)-(21) with n=200, 800, 3200 and p ≈ n^{1/2} (e.g., p=14, 28, 57). Record the total variation distance between the two posterior distributions of α and the frequentist coverage of 95% equal-tailed credible sets for each method. If the Algorithm 1 posterior does not converge to the same normal limit as the known-variance posterior, or if its coverage does not approach the nominal level, then Proposition 7 does not govern the proposed BDML procedure. An analytical counterpart would be to derive the marginal posterior of α under the inverse-Wishart prior by Laplace expansion and verify whether the extra O(p²/n²) and O(1/n) terms vanish fast enough to yield the BvM limit at p=o(√n).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central frequentist guarantees advertised in the abstract—asymptotic normality, semiparametric efficiency, and robustness to misspecification—rest on Proposition 7. That proposition explicitly assumes the true error variances σ2*ε and σ2*V are known, and places a prior directly on α. But Algorithm 1 and the simulations estimate α as the posterior ratio Σ12/Σ22 from a full inverse-Wishart (or LKJ/Cauchy) posterior on the reduced-form error covariance Σ; the error variances are not known. Proposition 7 therefore applies to a different posterior than the one the paper actually proposes, and no theorem in the paper bridges this gap. The proof sketch merely says the result follows from Walker (2025) after verifying his assumptions, and Remark 7 acknowledges a change from sup-norm to empirical L2 neighborhoods, asserting without demonstration that 'all proofs in Walker (2024) can be adapted.' The appendix checks some empirical-process bounds, but it does not derive the posterior of the ratio α=Σ12/Σ22 from the inverse-Wishart posterior, nor does it show that the ratio posterior has the claimed normal limit with variance 1/(n In(γ*)). Since the frequentist coverage of BDML credible intervals and the semiparametric efficiency claim are exactly what Proposition 7 is supposed to deliver, the central claim is not fully supported as stated. The abstract also promises empirical examples, but Section 6 contains only simulations; no empirical application, code, or data are provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Bayesian Double Machine Learning (BDML) for the partially linear model with high-dimensional controls. Instead of a structural regression of Y on (D, X), the method specifies a bivariate reduced-form regression of (Y, D) on X and recovers the causal parameter as the posterior ratio α = Σ12/Σ22 from the error covariance matrix. The authors argue that this reparameterization avoids regularization-induced confounding, that the implied prior on selection bias is non-dogmatic, and that the resulting estimator matches frequentist double machine learning in asymptotic bias (p²/n² versus p/n for a naive ridge/Bayes estimator) while improving finite-sample RMSE and coverage. The paper also claims a Bernstein-von Mises theorem, asymptotic normality, semiparametric efficiency, and robustness to misspecification, and it reports a simulation study comparing seven estimators.","tokens_in":30480,"tokens_out":3542,"duration_ms":41568,"significance":"If the advertised properties hold, BDML is a valuable and simple contribution: it gives applied researchers a fully Bayesian, conjugate-implementable procedure that avoids the tuning and cross-fitting sensitivities of frequentist DML while retaining its first-order guarantees. The core parameterization is clean, the posterior algebra in Section 4 is transparent, the distinction between naive shrinkage priors and reduced-form priors is well explained, and the simulation design is informative and clearly reported. The paper also gives explicit, falsifiable asymptotic rate claims (p/n versus p²/n²) that are directly comparable to the DML literature. However, the most consequential advertised guarantees—frequentist validity of credible intervals and semiparametric efficiency—are not actually proven for the algorithm as implemented, because the Bernstein-von Mises theorem is stated for a different posterior (known error variances) than the one used in Algorithm 1 and in the simulations.","major_comments":[{"comment":"Proposition 7 assumes that σ²*ε and σ²*V are known and places a prior directly on α, whereas Algorithm 1 and the Section 6 simulations estimate the full covariance matrix Σ (via inverse-Wishart or LKJ/Cauchy priors) and form α as the ratio Σ12/Σ22. No theorem in the paper derives the posterior distribution of that ratio or shows it has the normal limit stated in Proposition 7. The proof sketch only says that Walker's assumptions are verified, and Remark 7 asserts without demonstration that Walker's proofs can be adapted from sup-norm to empirical L2 neighborhoods. The appendix verifies some empirical-process bounds, but it does not bridge the gap between the known-variance posterior and the unknown-Σ ratio posterior. Since the abstract's claims of asymptotic normality, semiparametric efficiency, and 'valid frequentist confidence intervals' rest on this proposition, the central frequentist guarantee is not established for the method as proposed.","section":"Section 5.4, Proposition 7"},{"comment":"The derivation of the posterior bias of BDML relies on several uncontrolled approximations: the expansion |I−A|≈1−tr(A), a first-order Picard iteration for the posterior mode, and the identification of the posterior mean of Σ with the posterior mode plus a term of order O(1/n). Proposition 5 then asserts the p²/n² bias claim. These steps are heuristic rather than rigorous, and the claimed order of the bias, which is a headline result, is therefore not a theorem. The authors should either provide a complete proof with explicit remainders under Assumptions 2–5 or clearly state Proposition 5 as a heuristic expansion.","section":"Section 5.3, Propositions 4–5 and Appendix A"},{"comment":"Even taking the known-variance setup at face value, the adaptation of Walker (2025) to the empirical L2 norm is not established. The sentence 'All proofs in Walker (2024) can be adapted' is not a proof, and the paper does not state the precise conditions on the prior, the function class, or the metric that would make Walker's theorems apply verbatim. Given that this adaptation is load-bearing for the efficiency claim, the authors should provide a self-contained statement of which Walker assumptions hold and where the empirical-L2 modification enters each verification.","section":"Section 5.4, Remark 7 and proof of Proposition 7"}],"minor_comments":[{"comment":"The sentence beginning 'unless we specify a prior in which β and γ' is incomplete and garbled: it reads 'Treating σ2ε as known' mid-sentence. The intended logical point about Bayesian ignorability should be restated.","section":"Section 3, paragraph after Eq. (11)"},{"comment":"The abstract promises 'a number of empirical examples drawn from the recent literature,' but Section 6 contains only simulations. Either add the empirical applications or remove the promise from the abstract.","section":"Abstract and Section 6"},{"comment":"In the second display of Eq. (14), 'Var(Ui) = Var(εi + αVi) = Var(Ui) + α²Var(Vi)' should read 'Var(εi) + α²Var(Vi)'.","section":"Equation (14)"},{"comment":"Remark 2 says that 'if C̃n is negative definite, then X follows a multivariate Generalized Inverse Gaussian distribution,' but the subject should be Σ, not the data matrix X.","section":"Remark 2 and Appendix A"},{"comment":"The paper cites Walker (2025) in the introduction and references, but the proof of Proposition 7 and Remark 7 repeatedly refer to 'Walker (2024)'. The citation should be made consistent.","section":"References and proof of Proposition 7"},{"comment":"There are typographical errors in the captions: 'T able 1' in the table caption and 'labled' in the figure caption.","section":"Figure 1 and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The BvM gap is substantial and directly affects the abstract's headline claims. If the authors cannot close the gap, they should substantially temper the claims of frequentist validity and semiparametric efficiency, perhaps presenting the BvM result only for a stylized known-variance version. The paper's core reparameterization and finite-sample evidence are promising enough that the gap is fixable within the manuscript's scope, but as it stands the abstract overstates what is proved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the core idea is genuinely good. Estimate α from the covariance ratio of a reduced-form SUR, so you get DML-style robustness without the two-step plug-in. The algebra in Section 4 is correct, and the bias-rate comparison (p²/n² vs p/n) is a real addition to the Bayesian causal inference literature. But the headline guarantee—Bernstein–von Mises, semiparametric efficiency, robustness to misspecification—is proven only for a posterior in which the error variances are known and a prior is placed on α directly. Algorithm 1 and the simulations use an inverse-Wishart/LKJ posterior on the full error covariance and compute α as Σ12/Σ22. Those are different posteriors. The paper does not bridge that gap; it says all proofs in Walker (2025) can be adapted, which is an assertion, not a demonstration. That is the load-bearing soft spot, and it explains why the abstract overclaims.\n\nWhat’s new and good: the reframing is clean, and the observation that independent priors on reduced-form coefficients do not imply dogmatic selection-bias priors is correct and worth stating. The asymptotic expansion for the posterior mean of α (Proposition 5) is plausible, and Remark 4’s comparison with FDML is informative. The simulation study is honest: it includes the important competitors (Hahn et al., Linero, FDML full and split), and the main qualitative results—BDML-Hier covers well and gives short intervals—are believable. The paper is also well situated in the literature; the citation pattern is fine, and the external BvM source is cited. This is not a throwaway.\n\nWhere it’s soft: the known-variance assumption in Proposition 7 is not a minor technicality. The whole selling point is that BDML accounts for all sources of uncertainty by marginalizing over Σ. If the BvM result doesn’t cover the implemented posterior, then the frequentist validity of the credible intervals and the efficiency claim are not established. The proof sketch doesn’t even attempt to derive the posterior of the ratio from inverse-Wishart. Also, the abstract promises empirical examples, but Section 6 is only simulations. No code or data are provided, which limits reproducibility beyond the reported tables.\n\nWho this is for: applied microeconomists and Bayesians working on high-dimensional causal inference. They will get a useful method and a clear framework, even if the theoretical guarantees need tightening. It deserves a serious referee. I’d send it to peer review with the expectation that Proposition 7 is either made to cover the implemented posterior or the claims are revised downward. The method itself is worth publishing.","headline":"A clean and useful re-parameterization of DML with honest simulations, but the headline BvM/efficiency claims apply to a different posterior than the one implemented.","tokens_in":30998,"tokens_out":3233,"would_cite":true,"duration_ms":34711,"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":"Bayesian Double Machine Learning recovers the causal effect of a treatment from the covariance of reduced-form errors, avoiding regularization-induced confounding while matching the large-sample guarantees of frequentist double machine…","keywords":["causal inference","regularization-induced confounding","Bayesian methods","double machine learning","partially linear model","Bernstein-von Mises theorem","high-dimensional regression","shrinkage priors"],"falsifier":"Set known error variances, sub-Gaussian controls, and $p$ just below $\\sqrt{n}$ (e.g., $n=400$, $p=19$), simulate many datasets with $\\gamma^{*\\prime}\\beta^* \\neq 0$, and measure the frequentist coverage of BDML's 95% credible interval; if coverage is systematically below nominal, the asserted Bernstein–von Mises theorem for the empirical $L^2$ norm is not valid. Separately, at $n=400,p=200$ versus $n=800,p=565$, compare BDML and naive ridge empirical bias: the paper predicts BDML bias decays like $p^2/n^2$ and the naive bias like $p/n$.","tokens_in":29850,"feed_emoji":"🎯","tokens_out":16830,"duration_ms":153151,"temperature":0.7,"pith_summary":"The paper proposes and defends a fully Bayesian estimator, BDML, for the causal effect of a treatment $D$ on an outcome $Y$ in a partially linear model with many controls $X$: instead of shrinking the coefficient on $X$ in the structural regression $Y = \\alpha D + X'\\beta + \\varepsilon$, it fits a bivariate reduced-form regression of $(Y,D)$ on $X$ and reads $\\alpha$ from the error covariance, $\\alpha = \\sigma_{UV}/\\sigma_V^2$. The authors argue this sidesteps regularization-induced confounding, the bias that arises when regularized estimators only partially adjust for controls, because independent shrinkage priors on the reduced-form coefficients do not imply a dogmatic prior of zero selection bias. They claim BDML has asymptotic bias of order $p^2/n^2$ rather than $p/n$, is $\\sqrt{n}$-consistent when $p = o(n^{3/4})$, and satisfies a Bernstein–von Mises theorem that turns posterior credible sets into valid frequentist confidence intervals with semiparametric efficiency. If these claims hold, applied researchers get a Bayesian method that matches frequentist double machine learning asymptotically while giving better finite-sample uncertainty quantification. Simulations with $n=200$, $p=100$ show BDML with hierarchical shrinkage achieving near-nominal coverage, the lowest RMSE among competitors, and the shortest intervals among well-calibrated methods.","feed_headline":"Causal effect recovered from error covariance alone","feed_subtitle":"Bayesian DML matches frequentist double machine learning asymptotics and improves finite-sample coverage","key_machinery":"The load-bearing object is the reduced-form bivariate regression of $(Y,D)$ on $X$, a pair of regression equations whose errors are allowed to correlate (a seemingly unrelated regression). Its error covariance is $\\Sigma = \\begin{pmatrix} \\sigma_U^2 & \\sigma_{UV} \\\\ \\sigma_{UV} & \\sigma_V^2 \\end{pmatrix}$, and the causal effect is the ratio transform $\\alpha = \\sigma_{UV}/\\sigma_V^2$ applied to posterior draws of $\\Sigma$. This transform converts shrinkage from a problem about structural coefficients into a covariance-estimation problem: the reduced-form likelihood does not factorize, so independent priors on $\\delta$ and $\\gamma$ do not concentrate the prior on zero selection bias. Asymptotically, the posterior for $\\Sigma$ is approximately Inverse-Wishart with an exponential tilting term $\\exp(\\mathrm{tr}(\\tilde C_n \\Sigma)/2)$; the tilting matrix encodes the shrinkage-induced shift that produces the $p^2/n^2$ bias, and the Bernstein–von Mises theorem verifies that the induced posterior of $\\alpha$ is asymptotically normal around the true value.","core_discovery":"The central discovery is that the causal parameter of a partially linear model is identified by the covariance of the errors of two reduced-form regressions, and that a Bayesian posterior over that covariance yields a posterior over the causal effect directly. Writing $Y_i = X_i'\\delta + U_i$ and $D_i = X_i'\\gamma + V_i$, the structural assumption $\\mathrm{Cov}(\\varepsilon_i, V_i)=0$ implies $\\mathrm{Cov}(U_i,V_i)=\\alpha \\mathrm{Var}(V_i)$, hence $\\alpha = \\sigma_{UV}/\\sigma_V^2$. This reparameterization changes what shrinkage does: priors on $\\delta$ and $\\gamma$ induce a dependent prior on $\\beta = \\delta - \\alpha\\gamma$, so the model is not Bayesian-ignorable and the prior on selection bias stays non-degenerate as $p$ grows. The paper shows the posterior mean of $\\alpha$ is consistent, $\\sqrt{n}$-consistent under $p = o(n^{3/4})$, has asymptotic bias of order $p^2/n^2$, and shares a common asymptotic variance with frequentist double machine learning and the naive estimator. Under a Bernstein–von Mises theorem stated for known error variances, the posterior of $\\alpha$ is asymptotically normal, which makes BDML credible intervals asymptotically valid confidence intervals and establishes semiparametric efficiency.","pith_inferences":["Editorial extension: the identification $\\alpha = \\sigma_{UV}/\\sigma_V^2$ suggests a general template for Bayesian double debiasing—any causal parameter expressible as a ratio of reduced-form covariances, such as an instrumental-variables local average treatment effect, could be handled by the same covariance-posterior device. The paper does not develop this extension.","The bias calculations imply a practical calibration rule: prior precisions should be proportional to $p$ to keep shrinkage bias at order $p^2/n^2$. A reader could test whether BDML's advantage over the naive estimator grows with $p/n$ beyond the $1/2$ ratio used in the simulations.","The selection-bias result yields a diagnostic applicable before estimation: on a given dataset, compute how concentrated the naive prior is on zero selection bias relative to the BDML prior; the choice of method should matter most in designs with strong confounding, a prediction that could be checked by reweighting the simulation design.","The Bernstein–von Mises proposition assumes known error variances $\\sigma_\\varepsilon^2$ and $\\sigma_V^2$; until a version that integrates out these variances is proved, the semiparametric efficiency claim should be read as conditional on that assumption."],"forward_implications":["BDML is $\\sqrt{n}$-consistent whenever $p = o(n^{3/4})$, while the naive ridge-style estimator requires $p/\\sqrt{n} \\to 0$; for growth of $p$ between $\\sqrt{n}$ and $n^{3/4}$, the naive estimator is not $\\sqrt{n}$-consistent and BDML is.","BDML and FDML share an asymptotic bias of order $p^2/n^2$ and the same asymptotic variance, strictly beating the naive estimator's $p/n$ bias; consequently BDML dominates on asymptotic mean squared error.","Under the Bernstein–von Mises conditions, BDML's equal-tailed credible sets are asymptotically valid frequentist confidence intervals and the estimator attains semiparametric efficiency, so Bayesian inference can be reported with frequentist guarantees.","In the paper's simulation study, BDML with hierarchical shrinkage achieves coverage near 0.94, the lowest RMSE among all seven methods, and the shortest average interval width among methods with near-nominal coverage, across all values of the structural error variance $\\sigma_\\varepsilon$ studied.","The generative formulation allows hierarchical priors to adapt shrinkage to the data; the hierarchical version, which lets the variances of $\\delta$ and $\\gamma$ differ, is the best-performing method in the simulations."],"supporting_citations":[{"why":"Defines frequentist double machine learning, the estimator and asymptotic benchmarks that BDML aims to match.","marker":"Chernozhukov et al. (2018)"},{"why":"Names and analyzes regularization-induced confounding, giving the bias formula that motivates the BDML construction.","marker":"Hahn et al. (2018)"},{"why":"Introduces Bayesian ignorability, showing independent structural-coefficient priors are dogmatic about selection bias; its simulation design is used in Section 6.","marker":"Linero (2023)"},{"why":"Provides the semiparametric Bernstein–von Mises theorem for the partially linear model that Proposition 7 adapts.","marker":"Walker (2025)"},{"why":"Treats the partially linear model and the post-double-selection approach that supplies the rate conditions and confidence-interval construction used for FDML.","marker":"Belloni et al. (2014)"},{"why":"Presents the two-step Bayesian plug-in approach (HCPH) that BDML competes with and that motivates full marginalization over the first-stage parameters.","marker":"Hahn et al. (2020)"},{"why":"Supplies the R2D2 prior idea behind the hierarchical shrinkage version of BDML.","marker":"Zhang et al. (2022)"},{"why":"Documents the fragility of sparse selection, justifying the use of ridge rather than LASSO in the FDML first step.","marker":"Kolesár et al. (2025)"}],"fun_headline_variants":["Causal effect from reduced-form error covariance","Bayesian DML improves finite-sample causal inference","Bayesian double ML: lower bias, better coverage","Causal effect from covariance, not selection","Bayesian DML: efficient causal inference without selection bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything about Bayesian credible intervals being valid frequentist confidence intervals depends on an asserted but unproved adaptation of an existing Bernstein–von Mises theorem to the empirical $L^2$ norm under known error variances, so if that adaptation fails the inference guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["Causal effect from reduced-form error covariance","Bayesian DML improves finite-sample causal inference","Bayesian double ML: lower bias, better coverage","Causal effect from covariance, not selection","Bayesian DML: efficient causal inference without selection bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1484,"prompt_tokens":1012,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":628,"tokens_out":472,"duration_ms":4874,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:19:00.944124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set known error variances, sub-Gaussian controls, and $p$ just below $\\sqrt{n}$ (e.g., $n=400$, $p=19$), simulate many datasets with $\\gamma^{*\\prime}\\beta^* \\neq 0$, and measure the frequentist coverage of BDML's 95% credible interval; if coverage is systematically below nominal, the asserted Bernstein–von Mises theorem for the empirical $L^2$ norm is not valid. Separately, at $n=400,p=200$ versus $n=800,p=565$, compare BDML and naive ridge empirical bias: the paper predicts BDML bias decays like $p^2/n^2$ and the naive bias like $p/n$.","supporting_citations":[{"cited_title":", author Carvalho, C.M","cited_arxiv_id":null,"evidence_quote":"Names and analyzes regularization-induced confounding, giving the bias formula that motivates the BDML construction."},{"cited_title":"Parametrization, Prior Independence, and the Semiparametric Bernstein-von Mises Theorem for the Partially Linear Model","cited_arxiv_id":"2306.03816","evidence_quote":"Provides the semiparametric Bernstein–von Mises theorem for the partially linear model that Proposition 7 adapts."},{"cited_title":", author Murray, J.S","cited_arxiv_id":null,"evidence_quote":"Presents the two-step Bayesian plug-in approach (HCPH) that BDML competes with and that motivates full marginalization over the first-stage parameters."}],"review_version":2}