{"id":"9585fe62-2e00-4d8a-a930-c743ee27a662","arxiv_id":"2411.17910","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new Bayesian variable selection method for mediation analysis exploits correlations among high-dimensional mediators to improve detection of active exposure-mediator-outcome pathways.","lead":"This paper introduces a Bayesian method that finds which blood metabolites carry the effect of a Mediterranean diet on cardiometabolic risk, using correlations among metabolites to improve detection. It provides public code, simulation benchmarks, and a real-data illustration on two cohort substudies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported power gain of MRF-SSB over MVN-IB-SSB may be an artifact of per-dataset tuning of η, not of the MRF correlation structure.","rationale":"The reader's weakest_assumption centers on causal identification via the Imai-Yamamoto independent-mediators assumption. That is a genuine limitation for the real-data interpretation, and the paper acknowledges it. However, the paper's headline contribution is the simulation-based claim that MRF-SSB priors improve power to detect active pathways. That claim rests on a comparison in which only the proposed method receives per-dataset tuning of η through the phase-transition algorithm. Because η is chosen to increase the number of selected exposure-mediator pathways, the TPR gain relative to a fixed-prior IB model may be partly or wholly a tuning artifact. The reader's rationale did mention data-dependent tuning of η, but did not identify it as the weakest assumption; hence partial agreement. A concrete rerun with matched prior inclusion behavior or a correlation-blind calibration would settle whether the MRF's correlation structure is doing the work. If the gain disappears, the central claim should be substantially weakened; if the gain persists, the current conditional acceptance is appropriate. Thus the reader's CONDITIONAL verdict remains appropriate, but for a more specific and testable reason.","tokens_in":18213,"tokens_out":5970,"duration_ms":56532,"concrete_test":"Re-run Scenario I (and Scenario IV) with three additional arms: (1) MVN-IB-SSB with θγ increased so its average NVS(γ) matches MVN-MRF-SSB's NVS(γ); (2) MVN-IB-SSB with the same per-dataset η-calibration algorithm applied to a dummy correlation matrix with all pairwise correlations set to zero, so the calibration can raise inclusion without any true correlation signal; (3) MVN-MRF-SSB with η fixed at 0 (equivalent to IB) and with the tuned η, to isolate the contribution of η. If arm (1) or (2) reproduces most of the TPR(γ×ω) gap between MVN-IB-SSB and MVN-MRF-SSB, the paper's attribution of improved power to the MRF correlation structure is not supported. Report TPR, FPR, and NVS for both γ and γ×ω for all arms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (§4.3, Table 1) is that MRF-SSB priors increase power to detect active pathways by leveraging mediator correlations. But the comparison is not apples-to-apples. For MVN-MRF-SSB, η is chosen separately for each simulated dataset by the phase-transition algorithm in Section 3.2: the algorithm finds the largest grid value below the first η at which the posterior mean inclusion count exceeds the η=0 count by 0.05, so it is explicitly calibrated to increase sensitivity. MVN-IB-SSB is run with a fixed prior inclusion probability θγ = logit^-1(0.1). In Scenario I this yields NVS(γ)=24.5 versus 18.3 and TPR(γ)=68.7% versus 58.4%, but also FPR(γ)=1.4% versus 0.3%. The improved γ×ω FPR (0.7% versus 1.6%) is produced by the SSB outcome-stage filter, not by the MRF prior. If the same sensitivity boost were given to the IB prior—for example, by raising θγ, or by applying the same η-calibration algorithm to a correlation-blind model—a substantial part of the TPR gap might persist, directly contradicting the paper's attribution of the gain to the MRF's exploitation of mediator correlations. Scenario IV shows the same pattern. The causal-independence limitation identified by the reader is real but secondary for the power claim; the main claim is an empirical comparison that is currently confounded by asymmetric tuning.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian variable selection framework for high-dimensional mediation analysis with correlated mediators. The method combines a Markov random field (MRF) prior on the exposure-mediator selection indicators, which leverages mediator correlations, with sequential subsetting Bernoulli (SSB) priors on the mediator-outcome selection indicators, which encourage joint selection of active pathways. The authors present simulation studies comparing the proposed MVN-MRF-SSB model with MVN-IB-SSB and Normal-IB-SSB baselines across four scenarios, and apply the method to metabolomics data from the HPFS and NHSII cohorts to identify metabolites mediating the relationship between Mediterranean diet adherence and cardiometabolic risk. The central empirical claim is that the MRF prior improves power to detect active exposure-mediator-outcome pathways while maintaining specificity.","tokens_in":18550,"tokens_out":5714,"duration_ms":45193,"significance":"If the central claim holds, the proposed method would be a useful addition to the toolkit for high-dimensional mediation analysis in correlated omics data. The paper includes publicly available code, a realistic application, and a comparison against simpler variants of its own model. The methodological contribution is the integration of MRF and SSB priors in a joint mediation model, which is a sensible extension of existing Bayesian variable selection ideas. However, the strength of the evidence for the main claim is limited by the confounded comparison protocol and the omission of the most relevant existing competitor, so the practical significance is currently not fully established.","major_comments":[{"comment":"The comparison between MVN-MRF-SSB and MVN-IB-SSB is confounded by asymmetric prior calibration. The MRF prior's η is selected per simulated dataset by the phase-transition algorithm in Section 3.2, which by construction chooses the largest grid value below the first η at which the posterior mean inclusion count exceeds the η=0 count by 0.05. In contrast, the IB prior uses a fixed θγ = logit^{-1}(0.1) with no analogous calibration. In Scenario I this yields NVS(γ)=24.5 versus 18.3 and TPR(γ)=68.7% versus 58.4%, but also FPR(γ)=1.4% versus 0.3%. The observed power gain may therefore reflect the sensitivity boost from data-dependent tuning rather than the MRF's exploitation of mediator correlations. The authors should re-run the comparison with matched calibration—for example, choosing θγ for the IB prior to achieve the same expected model size, or applying the same phase-transition algorithm to a model with η=0—and report whether the TPR gap persists.","section":"Sections 3.2 and 4.3, Table 1"},{"comment":"The simulation study omits the most relevant existing competitor. The Introduction cites Song et al. (2021a) as a joint mediation framework that handles correlated mediators through a latent neighboring structure, but the simulation comparisons include only Normal-IB-SSB and MVN-IB-SSB, which are simplifications of the proposed model rather than independent published methods. Without a comparison against Song et al. (2021a) or another existing joint mediation method, the abstract's claim of 'superior power' over existing approaches is unsupported. The authors should add such a comparison or restrict the claim to the comparison with the independent-Bernoulli baseline.","section":"Sections 4.1 and 4.2"},{"comment":"The claim that specificity is maintained is not supported at the exposure-mediator stage. In Scenario I, MVN-MRF-SSB increases FPR(γ) from 0.3% to 1.4%, more than a fourfold increase, and in Scenario IV from 0.3% to 0.7% with a standard deviation of 1.6. The lower FPR for the final γ×ω pathway (0.7% versus 1.6% in Scenario I) is produced by the SSB prior's pruning in the outcome model, not by the MRF prior. The text should distinguish the two stages and avoid describing the FPR increase as 'slight'; the tradeoff should be reported transparently.","section":"Section 4.3, Table 1"}],"minor_comments":[{"comment":"The statement that 'no existing statistical methods for joint mediation analysis leverage mediator correlations through specialized priors or penalty terms' is inconsistent with the later citation of Song et al. (2021a), which is described as employing a neighboring structure for correlated mediators; please reword to acknowledge prior work.","section":"Section 1"},{"comment":"The definition of γ(ηg) as the '50-th percentile of the posterior mean estimates of γ' is ambiguous; please clarify whether it is the median over mediators of the posterior inclusion probabilities or the median over MCMC iterations.","section":"Section 3.2"},{"comment":"Several reported 95% HPDIs for indirect effects include zero (e.g., C51:3 TG: 0.15 (0.00, 0.25); C22:6 LPC: -0.08 (-0.13, 0.00)), while these pathways are listed as selected; since selection is based on PPI thresholds, please add a note explaining how to interpret such intervals.","section":"Section 5.3, Table 3"},{"comment":"There is a typo in 'provides an detailed analysis'; it should be 'a detailed analysis'.","section":"Section 1, last paragraph"},{"comment":"The paper would benefit from a table or figure showing the distribution of the selected η values across the 120 simulated datasets, to help readers assess the variability in the data-dependent calibration.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the code is publicly available, which is a strength. The data-dependent η calibration is a concern beyond the fairness of the comparison: it uses the data to set the prior, so the reported posterior intervals for indirect effects and the FDR control may not have their nominal properties. If the authors cannot provide a matched-calibration comparison, the editor may wish to request a more conservative interpretation of the simulation results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate and mostly clear extension of MRF-based variable selection to high-dimensional mediation, with code and honest limitations, but the paper's headline power gain is confounded by the way eta is tuned for MRF and not for the Bernoulli comparator. Don't take Table 1's TPR gap at face value.\n\nWhat's new: the combination of an MRF prior on the exposure-mediator indicators and a sequential subsetting Bernoulli (SSB) prior on the mediator-outcome indicators, with the latter receiving feedback from the former. That's a sensible way to coordinate selection across the two models, and the cutting-feedback analogy is apt. The paper also ships code and runs a null scenario and a misspecified-covariance scenario, and it is upfront about the causal-independence assumption and the restrictiveness of the factor-analytic covariance. Those are real strengths.\n\nWhere it's soft: the main simulation comparison -- MVN-MRF-SSB versus MVN-IB-SSB -- is not apples-to-apples. For MRF, eta is chosen per dataset by the phase-transition algorithm in Section 3.2, which effectively looks for the eta just below the point at which the number of selected gammas jumps by 0.05. That is explicitly calibrated to raise sensitivity. The IB comparator gets a fixed theta_gamma = logit^-1(0.1). So the MRF's higher TPR (68.7% vs 58.4%) and also higher gamma FPR (1.4% vs 0.3%) may simply reflect that the MRF was tuned to select more. Matching the IB model's prior inclusion probability upward -- or applying the same tuning algorithm to a correlation-blind model -- would tell you how much of the gain is actually due to the MRF exploiting correlations. As it stands, the central claim is under-supported.\n\nAlso missing: Song et al. 2021a, which does coordinated selection of correlated mediators, is cited but not included in the simulations. Given the paper is claiming superiority over existing joint mediation methods, that omission is noticeable.\n\nMinor: the gamma x omega FPR improvement (0.7% vs 1.6%) is attributable to the SSB outcome filter rather than the MRF, and the paper should say so more clearly. The causal-independence assumption is acknowledged and is standard for this literature, so I don't weigh it heavily.\n\nOverall: worth a serious referee. The method is not obviously wrong, the writing is clear, and the code and simulations provide a base to build on. But the authors should be asked to redo the comparison with matched sensitivity, and to include the closest competitor, before the power claim is accepted.\n\nRecommendation: send to peer review with a request for these revisions.","headline":"Solid method paper with a real tuning confound in its headline comparison; the power gain is not yet convincingly attributed to the MRF prior.","tokens_in":19078,"tokens_out":3074,"would_cite":false,"duration_ms":25398,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62J05","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Modeling mediator correlations in the selection prior increases the power of high-dimensional mediation analysis to find active exposure–mediator–outcome pathways.","keywords":["Bayesian variable selection","high-dimensional mediation analysis","Markov random field prior","sequential subsetting Bernoulli prior","spike-and-slab prior","metabolomics","indirect effects","phase transition"],"falsifier":"Rerun the Scenario I comparison with the mediator indices randomly permuted in the MRF edge weights $c_{r,j}$—same mediator data, same correlations, but the prior's edges no longer point at the true correlation neighborhoods. If MVN-MRF-SSB still beats MVN-IB-SSB on the combined-pathway TPR, the correlation structure is not the source of the gain; if the advantage vanishes, the paper's mechanism is confirmed.","tokens_in":112,"feed_emoji":"🧪","tokens_out":8791,"duration_ms":135930,"temperature":0.7,"pith_summary":"The paper proposes a Bayesian variable-selection method for mediation analysis with many correlated mediators, the setting of metabolomics studies where hundreds of small molecules sit between an exposure and a health outcome. Its central claim is that modeling the mediator correlation structure in the prior for which pathways are active—rather than treating each mediator's inclusion independently—increases the chance of finding genuine exposure–mediator–outcome pathways without raising false positives. In simulations the proposed MVN-MRF-SSB model detected 68% of active pathways versus 57% for the independent-Bernoulli version when mediators were correlated, and it kept the false positive rate for indirect effects at 0.7% versus 1.6%. In two cohort substudies it identified 15 metabolites as mediators of the Mediterranean diet–cardiometabolic risk association, where the comparison model found 8, and it estimated indirect effects with uncertainty intervals for each selected pathway.","feed_headline":"Correlated mediators lift pathway detection from 57% to 68%","feed_subtitle":"Using mediator correlations in the prior finds 15 diet–heart pathways, versus 8 without.","key_machinery":"The load-bearing object is the MRF-SSB prior pair. The Markov random field prior for $\\gamma_j$, the exposure–mediator inclusion indicator, is $\\pi(\\gamma_j \\mid \\gamma_{(-j)}, \\Sigma) = \\mathrm{logit}^{-1}(\\theta_\\gamma + \\eta \\sum_{r \\neq j} |c_{r,j}| \\gamma_r)$, so a mediator's inclusion probability rises when correlated mediators are included, with $c_{r,j}$ taken from the mediator covariance. The sequential subsetting Bernoulli prior for $\\omega_j$, the mediator–outcome inclusion indicator, is $\\pi(\\omega_j) = \\tilde{\\gamma}_j \\theta_\\omega^{\\omega_j}(1-\\theta_\\omega)^{1-\\omega_j} + (1-\\tilde{\\gamma}_j) I_0$, where $\\tilde{\\gamma}_j$ is the realized exposure–mediator inclusion; this transmits mediator-model selections to the outcome model but not vice versa. These priors sit on spike-and-slab distributions for $\\tau_j$ and $\\delta_j$, with the mediator covariance estimated by a factor-analytic structure $\\Sigma = \\sigma_\\Sigma^2(\\lambda\\lambda^\\top + I_q)$, and the MRF coupling strength $\\eta$ is placed below a phase-transition boundary found by posterior simulation. The MRF part is what converts mediator correlations into extra detection power; the SSB part is what keeps the eventual indirect-effect selection sparse.","core_discovery":"The paper's central claim is that the standard practice of assigning independent Bernoulli priors to the binary selection indicators in high-dimensional mediation analysis discards useful information: mediators that are correlated with an active mediator are more likely to carry an exposure effect, and the proposed Markov random field prior exploits exactly that. Coupled with a sequential subsetting Bernoulli prior that feeds the selected exposure–mediator pathways forward into the mediator–outcome selection while cutting feedback in the reverse direction, the prior pair improves the true positive rate for detecting active pathways in correlated settings. In the paper's Scenario I the gain is 68.1% versus 56.9% TPR on the combined pathways, with a lower false positive rate for indirect effects (0.7% versus 1.6%). In the motivating metabolomics analysis the prior identifies seven additional diet-related metabolites not selected by the independent-prior model, among them fish-intake markers such as C22:6 lysophosphatidylcholine.","pith_inferences":["A testable extension the authors mention but do not develop is to replace the continuous mediator model with generalized linear models for binary or count mediators; the same MRF-SSB prior should carry over and the power gain can be checked in simulations.","Because the SSB prior cuts feedback from the outcome model back to exposure–mediator selection, the method is asymmetric by design: a mediator with a strong outcome association but no detected exposure association can never enter the final pathway set; analysts should decide whether that asymmetry fits their question.","The phase-transition calibration for $\\eta$ is a heuristic posterior-simulation step; an automatic data-driven rule would remove the most delicate tuning choice and would clarify how much of the reported gain depends on tuning rather than on the prior's structure."],"forward_implications":["In correlated-mediator settings, the proposed MVN-MRF-SSB model detects active exposure–mediator–outcome pathways at a higher true positive rate than the independent-Bernoulli counterpart (68.1% versus 56.9% in Scenario I).","The power gain disappears when mediators are uncorrelated (Scenario II), showing the improvement is attributable to the correlation structure rather than to generally more aggressive selection.","Under a null model with no active pathways (Scenario III), the proposed method selects essentially no pathways (mean NVS 0.4), so the correlation-aware prior does not inflate false discoveries in the absence of true effects.","The sequential subsetting prior keeps the final indirect-effect false positive rate low even when the MRF prior raises the candidate pool at the exposure–mediator stage (0.7% versus 1.6% in Scenario I).","In the metabolomics application the method reports 15 active mediators of the Mediterranean diet–cardiometabolic risk association, including additional fish-intake markers, with posterior-median indirect effects and 95% intervals."],"supporting_citations":[{"why":"Supplies the causally-independent-mediators assumption under which each indirect effect is nonparametrically identified.","marker":"Imai and Yamamoto (2013)"},{"why":"Introduces the Markov random field prior for structured variable selection and the phase-transition phenomenon that guides the choice of $\\eta$.","marker":"Li and Zhang (2010)"},{"why":"Provides the multivariate MRF-based variable selection approach and posterior-simulation phase-transition calibration that the paper adapts to mediation analysis.","marker":"Lee et al. (2017)"},{"why":"Contributes the spike-and-slab priors that underlie the stochastic search variable selection for $\\tau_j$ and $\\delta_j$.","marker":"George and McCulloch (1993, 1997)"},{"why":"Supplies the factor-analytic covariance model $\\Sigma = \\sigma_\\Sigma^2(\\lambda\\lambda^\\top + I_q)$ that makes high-dimensional posterior updates computationally feasible.","marker":"Hogan and Tchernis (2004)"},{"why":"Establishes the prior-simulation grid search over $\\eta$ to locate the phase-transition boundary used for hyperparameter specification.","marker":"Stingo et al. (2011)"},{"why":"Represents the existing joint mediation modeling approach with simple latent neighboring structure that the paper contrasts with its correlation-aware priors.","marker":"Song et al. (2021a)"},{"why":"Provides the adaptive Bayesian FDR thresholding method used to choose the posterior inclusion probability cutoff for variable selection.","marker":"Wadsworth et al. (2017)"}],"fun_headline_variants":["Correlated mediators lift pathway detection from 57% to 68%","Bayesian prior uses mediator correlations to find more active pathways","Mediator-correlation prior finds 7 additional diet-related metabolites","Detection rate jumps 11% with mediator-correlation prior","MRF prior improves mediation TPR from 57% to 68%"],"cache_read_input_tokens":21120,"weakest_assumption_plain":"The method's causal interpretation relies on the Imai–Yamamoto assumption that every mediator operates through an independent pathway, with no causal effect of one mediator on another and no unmeasured common cause of mediators and outcome; if mediators affect each other, the estimated indirect effects are not identified as causal.","fun_headline_variants_meta":{"raw":{"variants":["Correlated mediators lift pathway detection from 57% to 68%","Bayesian prior uses mediator correlations to find more active pathways","Mediator-correlation prior finds 7 additional diet-related metabolites","Detection rate jumps 11% with mediator-correlation prior","MRF prior improves mediation TPR from 57% to 68%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2338,"prompt_tokens":999,"completion_tokens":1339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1250}},"tokens_in":615,"tokens_out":1339,"duration_ms":9395,"temperature":1.0,"reasoning_tokens":1250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:41:54.636839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Scenario I comparison with the mediator indices randomly permuted in the MRF edge weights $c_{r,j}$—same mediator data, same correlations, but the prior's edges no longer point at the true correlation neighborhoods. If MVN-MRF-SSB still beats MVN-IB-SSB on the combined-pathway TPR, the correlation structure is not the source of the gain; if the advantage vanishes, the paper's mechanism is confirmed.","supporting_citations":[{"cited_title":"and Yamamoto, T","cited_arxiv_id":null,"evidence_quote":"Supplies the causally-independent-mediators assumption under which each indirect effect is nonparametrically identified."},{"cited_title":"H., Tadesse, M","cited_arxiv_id":null,"evidence_quote":"Provides the multivariate MRF-based variable selection approach and posterior-simulation phase-transition calibration that the paper adapts to mediation analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the spike-and-slab priors that underlie the stochastic search variable selection for $\\tau_j$ and $\\delta_j$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the factor-analytic covariance model $\\Sigma = \\sigma_\\Sigma^2(\\lambda\\lambda^\\top + I_q)$ that makes high-dimensional posterior updates computationally feasible."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the prior-simulation grid search over $\\eta$ to locate the phase-transition boundary used for hyperparameter specification."},{"cited_title":"D., Argiento, R., Guindani, M., Galloway-Pena, J., Shelburne, S","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive Bayesian FDR thresholding method used to choose the posterior inclusion probability cutoff for variable selection."}],"review_version":1}