{"id":"e58d11f1-fa61-4ec1-983b-cdb6b4371dda","arxiv_id":"2506.07953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A mediation analysis method for scalar exposures, high-dimensional mediators, and sparse irregular longitudinal outcomes identifies five lipid metabolites as potential mediators between rest-activity rhythms and cognitive decline in older men.","lead":"This paper proposes a statistical method for finding which of many lipid biomarkers carry the effect of sleep-wake rhythm on cognitive decline measured at few, unevenly spaced visits. The method is applied to MrOS Sleep Study data and identifies five lipid metabolites as candidate mediators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FDR control is unverified: the permutation p-value targets a marginal model, not the full mediation null, and no simulation reports empirical FDR, so the central claim rests on untested assumptions.","rationale":"The reader's weakest assumption (independence of Pα and Pβ) is one facet of the problem, but I find a more fundamental gap: the p-value for the mediator-to-outcome effect is obtained from a marginal model that omits other mediators, while the null and the NIE in Eq. (3) refer to the full model. With correlated mediators, the marginal coefficient is not the full-model coefficient, so the test may reject for non-mediators. The paper gives no argument that the marginal test controls size for the full-model null, and Theorem 1 is stated for a 'correctly specified time-varying model' while the estimation procedure uses a purposefully marginal model. In addition, no simulation evaluates empirical FDR, so the central methodological claim has no direct empirical support; the identical values in Table 1 across dense and sparse scenarios further undermine confidence in the numerical evidence. These issues are addressable—one could use joint-model p-values, clearly reinterpret the claim as FDR control for the marginal null, and add FDR simulations—so the paper should not be accepted as is, but the conditional verdict remains appropriate rather than outright rejection.","tokens_in":16242,"tokens_out":11356,"duration_ms":143299,"concrete_test":"Reproduce the simulation design of Section 5.1 (Scenarios 2 and 3, all four correlation cases) with p = 50 and the stated correlated Σϵ. Run the proposed procedure on each of the 100 replicates, using the true generative model to label each mediator as null or non-null according to the full-model condition α_k β_k(t) = 0 (with β_k(t) from Eq. (2), not the marginal model). Compute the empirical FDR of the selected set at nominal b = 0.05 and b = 0.10, with a 95% confidence interval. Also compute the rejection rate of the marginal permutation test for a mediator with α_k ≠ 0 but full-model β_k(t) = 0 that is correlated with a true mediator; if this rejection rate exceeds its nominal size, or if the empirical FDR exceeds the nominal level beyond sampling error, the central FDR-control claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is FDR control for the null H0: α_k = 0 or β_k(t) = 0, with β_k(t) the full-model coefficient in Eq. (2). However, Pβ_k(t) is computed from the marginal model Eq. (5), which omits all other mediators. When mediators are correlated (as in the simulation's Σϵ = 0.1^{|k1−k2|} and in the MrOS lipid panel), the marginal coefficient β_k^*(t) can be nonzero even when the full-model β_k(t) = 0, so the test is not valid for the stated null. The paper provides no theorem showing that the marginal and full-model coefficients coincide or that the FDR bound in Theorem 1 transfers to the actual marginal procedure. Theorem 1 also assumes, without evidence or sensitivity analysis, that Pα_k and Pβ_k(t) are asymptotically independent for each k. Critically, Section 5 never reports an empirical FDR for the proposed screening procedure; it reports power, screening frequencies of true mediators, and estimation error, but not false-discovery behavior. The central FDR-control claim is therefore supported only by a theorem whose assumptions are unverified and whose test statistic targets a different null. This matters directly for the application: the five consistently selected lipids are all fatty acids or acylcarnitines, so correlated marginal signals, rather than true mediation, could drive their selection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a mediation analysis method for scalar exposures, high-dimensional mediators, and sparse, irregularly spaced longitudinal outcomes. The exposure-to-mediator effects are tested with OLS p-values, and the mediator-to-outcome time-varying effects are tested with a permutation test applied to a marginal functional regression model. The two p-values are combined through the HDMT JS-mixture procedure, with a theorem claiming asymptotic control of the estimated FDR. The method is evaluated in simulations under four correlation structures and three sampling designs, and applied to the MrOS Sleep Study to identify lipid metabolites mediating the association between acrophase and cognitive decline, with five metabolites consistently selected.","tokens_in":16517,"tokens_out":3637,"duration_ms":47593,"significance":"If the central FDR-control claim is valid, the paper would fill a genuine gap: high-dimensional mediation analysis with functional or longitudinal outcomes that are measured sparsely and irregularly. The application to a real cohort with 476 lipid metabolites is timely and substantively interesting, and the idea of combining a permutation test with the HDMT mixture FDR is a sensible starting point. The paper also gives explicit credit to existing work on functional regression and high-dimensional mediation. However, the central theoretical guarantee is currently unverified: the theorem relies on an unproved independence assumption and on a supplement that is not available, and the simulation study does not report empirical FDR. These issues are load-bearing for the paper's main claim.","major_comments":[{"comment":"The permutation p-value P_βk(t) is computed from the marginal model (5), which omits all other mediators, whereas the null H0 in Eq. (4) is stated in terms of β_k(t) from the full model (2). When mediators are correlated—as in the simulation's Σ_ϵ = 0.1^{|k1−k2|} and in the MrOS lipid panel—the marginal coefficient β_k^*(t) can be nonzero even when the full-model β_k(t) is zero. The paper provides no theorem showing that marginal and full-model coefficients coincide, nor any simulation demonstrating FDR control under this mismatch. This is a direct threat to the validity of the screening procedure for the stated mediation null.","section":"Section 4.1, Eq. (5) vs. Eq. (2)"},{"comment":"Theorem 1 assumes that P_αk and P_βk(t) are asymptotically independent for each k, but this assumption is asserted without proof, reference, or sensitivity analysis. The proof is said to be in Section S1 of a supplementary file that is not available to the reader. Moreover, the theorem only bounds the estimated quantity FDR-hat(λ), not the true FDR, so even if the theorem is correct it does not establish control of the actual false discovery rate. The authors should either provide the proof and a justification of independence, or add simulation evidence on empirical FDR under correlated mediators.","section":"Theorem 1, Section 4.1"},{"comment":"No simulation result reports an empirical FDR or false-positive count for the proposed screening procedure. Section 5.4 reports screening frequencies for the true mediators only, and Section 5.3 reports power for one mediator. Without a direct report of FDR or type I error, the paper's central claim of false discovery control is not supported by the numerical study. I request an explicit FDR table or figure under the null and under the full simulation design.","section":"Section 5.3–5.4"},{"comment":"The mediation effect estimation is evaluated only under oracle knowledge of the true mediators: the text states that 'the first p0 = 4 mediators were correctly identified in each simulation replicate.' This does not reflect the actual pipeline, where screening errors propagate into the estimation of direct and indirect effects. Furthermore, within each case in Table 1, the reported bias and standard deviation are identical across Scenarios 1, 2, and 3, which is surprising given that the sampling design changes the number of repeated measurements per subject; this needs clarification or correction.","section":"Section 5.5 and Table 1"}],"minor_comments":[{"comment":"The phrase '5 consequtive 24-hour periods' contains a typo; it should be 'consecutive.'","section":"Section 2"},{"comment":"The definition of the overall p-value P_k = max(P_αk, P_βk(t)) should be reconciled with the three-component mixture decomposition; the authors should state explicitly how the mixture proportions are identifiable under the asymptotic independence assumption.","section":"Section 4.1, Eq. (4)"},{"comment":"The notation is inconsistent: the text uses bλb for the threshold and then writes FDR-hat(λ); please define all symbols clearly and use a uniform notation for the estimated FDR and the threshold.","section":"Section 4.1"},{"comment":"The condition that 'p is bounded by a polynomial function of L_n' in Theorem 1 is vague; since the theorem takes p → ∞, the relationship between p and L_n should be specified.","section":"Section 4.1"},{"comment":"The proof of Theorem 1 is deferred to a supplementary file that is not included with the arXiv version; for a journal submission, the supplement should be made available to reviewers.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a genuine applied motivation and a plausible algorithmic pipeline, but the central FDR-control claim is currently supported by neither a verifiable proof nor direct simulation evidence. The mismatch between the marginal testing model and the full-model null hypothesis is a substantive concern that may require more than a local fix; however, it is conceivable that the authors can resolve it by redefining the null, providing a new theorem, and adding empirical FDR simulations. I therefore recommend major revision rather than rejection, but the revision must address the marginal-versus-full-model issue directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new part is real: nobody else has combined high-dimensional mediators with sparse, irregularly spaced longitudinal outcomes using spline-based time-varying coefficients, a permutation F-statistic, and JS-mixture FDR control. That combination is not in Zeng et al. (2021), which handles sparse longitudinal data only in a low-dimensional mediation setting. The MrOS application is a sensible, substantive test case, and the consistent identification of five lipid metabolites at FDR 0.05 is biologically coherent.\n\nThe execution, though, has soft spots that matter. The central FDR control claim is not empirically checked: the simulations report screening frequencies of true mediators and marginal power for one coefficient, but no empirical FDR for the full screening procedure. That is a conspicuous omission for a paper whose headline result is asymptotic FDR control. The permutation test targets the marginal model (5) for each mediator, while the stated null is about the full-model coefficient beta_k(t) in (2). With correlated mediators, the marginal coefficient can be nonzero even when the full-model coefficient is zero, so the p-value may not be valid for the mediation null. Theorem 1 assumes asymptotic independence between P_alpha_k and P_beta_k(t) without sensitivity analysis, and the proof is relegated to a supplementary file that is not available. The estimation evaluation assumes the first four mediators are correctly identified, so it does not reflect the full screening-plus-estimation pipeline. The identical values across scenarios in Table 1 look like a reporting artifact and should be clarified. There is also no comparison against Zeng et al. (2021), the closest prior, and no code or data.\n\nThese are addressable rather than fatal. The paper is a plausible extension, clearly written, and the authors appear to have thought carefully about the correlation structures and the sparse-data problem. It is not ready in this form, but the core idea deserves referee time.\n\nIf I were editing, I would send it to peer review with a specific request: add simulations reporting empirical FDR for the proposed procedure, address the marginal-versus-full model mismatch, prove or weaken the independence assumption, add a comparison with Zeng et al., and provide code and data. That would turn a conditional contribution into a more solid one.","headline":"Genuinely new combination of high-dimensional mediation with sparse irregular longitudinal outcomes, but the FDR control claim is unverified and the permutation test targets a different null.","tokens_in":17031,"tokens_out":1840,"would_cite":false,"duration_ms":23489,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A permutation test plus a three-component mixture FDR controls mediation discoveries in sparse longitudinal outcomes and identifies five lipid mediators linking sleep-wake timing to cognitive decline.","keywords":["mediation analysis","high-dimensional mediation","sparse longitudinal data","time-varying coefficient models","permutation test","false discovery rate","rest-activity rhythms","cognitive decline"],"falsifier":"Simulate null data with no true mediation but with shared subject-level random effects that couple the mediator and outcome equations, so $P_{\\alpha_k}$ and $P_{\\beta_k(t)}$ are dependent, and run the proposed permutation and mixture-FDR procedure with growing $p$; if the empirical FDR exceeds the target level as $p$ grows, the independence assumption behind Theorem 1 is violated.","tokens_in":16025,"feed_emoji":"🧠","tokens_out":14706,"duration_ms":133459,"temperature":0.7,"pith_summary":"Mediation studies usually break down when the outcome is measured only a few times per person at uneven intervals and hundreds of candidate mediators are tested at once. This paper proposes a pipeline that models each person's outcome trajectory with splines, accounts for within-subject correlation through a weighted least squares fit, tests each mediator's time-varying effect with a permutation F-test, and then feeds the two per-mediator p-values into a three-component mixture false discovery rate procedure. The claim is that this pipeline controls the FDR asymptotically even when the longitudinal data are sparse, and that it outperforms a pointwise functional-regression alternative in simulations. In the MrOS Sleep Study, the pipeline consistently selects five lipid metabolites at an FDR of 0.05 as mediators of the relationship between acrophase—the time of daily peak activity—and cognitive decline, whereas the comparison method selects none. A sympathetic reader would take this as evidence that careful trajectory modeling plus a mixture FDR formula can turn sparse follow-up data into usable mediation discoveries.","feed_headline":"Five lipids mediate the sleep-rhythm link to cognitive decline","feed_subtitle":"A mixture-FDR mediation test for sparse, irregular visits pinpoints fatty-acid metabolites in older men.","key_machinery":"The load-bearing object is the three-component mixture FDR estimator for the overall mediation p-value. Each mediator's overall p-value is $P_k = \\max(P_{\\alpha_k}, P_{\\beta_k(t)})$, so under the composite null ($\\alpha_k = 0$ or $\\beta_k(t) = 0$) it falls into one of three components; the FDR numerator is $\\hat{\\pi}_{01}\\lambda + \\hat{\\pi}_{10}\\lambda + \\hat{\\pi}_{00}\\lambda^2$, and the threshold is the largest $\\lambda$ with estimated FDR below the target. The other half of the machinery is the permutation test for $\\beta_k(t)$: the outcome is projected onto a B-spline basis, the model is fit by weighted least squares with an estimated within-subject correlation matrix, and the mediator's importance is scored by the F-statistic comparing weighted residual sums of squares with and without the mediator. This test is what supplies a valid $P_{\\beta_k(t)}$ under sparse and irregular sampling, and the mixture formula is what converts 476 candidate p-values into a controlled discovery set.","core_discovery":"On its own terms, the paper's central discovery is that high-dimensional mediation testing can be made to work when the outcome is a sparse, irregularly sampled longitudinal process. For each candidate mediator $k$, the exposure-to-mediator effect $\\alpha_k$ is tested by ordinary least squares, and the mediator-to-outcome time-varying effect $\\beta_k(t)$ is tested by permuting the mediator values and comparing weighted residual sums of squares from a B-spline expansion of the marginal model; the overall mediation p-value is $P_k = \\max(P_{\\alpha_k}, P_{\\beta_k(t)})$. Under the null, $P_k$ is modeled as a three-component mixture, and the estimated FDR at threshold $\\lambda$ is $(\\hat{\\pi}_{01}\\lambda + \\hat{\\pi}_{10}\\lambda + \\hat{\\pi}_{00}\\lambda^2) / (\\max\\{1, R(\\lambda)\\}/p)$. Theorem 1 states that, for a correctly specified time-varying model with asymptotically independent $P_{\\alpha_k}$ and $P_{\\beta_k(t)}$, sufficiently many permutations, and $p$ growing at most polynomially in the number of basis functions, the estimated FDR controls the target level asymptotically: $\\limsup_{p\\to\\infty} \\widehat{\\mathrm{FDR}}(\\lambda) \\le b$. Applied at FDR 0.05 to 476 lipid mediators in 490 older men, the procedure consistently selects five metabolites—3-hydroxymyristate, 3-hydroxylaurate, tetradecadienoate (14:2), dodecadienoate (12:2), and 3-hydroxyoctanoyl-carnitine—as mediators linking acrophase (the time of daily peak activity) to cognitive decline, while the pointwise function-on-scalar regression comparator identifies none.","pith_inferences":["One testable extension is to check robustness of the MrOS list to the stated independence assumption: generate null data with shared subject-level random effects that induce dependence between $\\hat{\\alpha}_k$ and the permutation statistic for $\\beta_k(t)$, and compare the empirical FDR with the nominal level.","The procedure could be adapted to other actigraphy summary measures, such as amplitude or relative amplitude, since the mediation model itself is exposure-agnostic.","If the five lipids replicate in an independent cohort, the estimated time-varying indirect effects shown in the paper could be used to identify the follow-up window in which the lipid pathway matters most for cognitive decline."],"forward_implications":["If the FDR control holds, the same spline-plus-permutation pipeline can be used for any scalar exposure, high-dimensional mediator panel, and sparse longitudinal outcome, not just cognition and lipids.","The MrOS finding implies that fatty-acid and acylcarnitine metabolism deserves closer study as a pathway from rest-activity rhythm disruption to cognitive decline in older men.","The permutation test for time-varying coefficients is itself portable and can be used to test functional predictors in longitudinal regression outside the mediation setting.","Because the five mediators are selected consistently across all four user-specified correlation structures at FDR 0.05, the authors show the discovery does not depend on the chosen covariance model, although the larger set at FDR 0.10 does vary with it."],"supporting_citations":[{"why":"Supplies the three-component mixture FDR procedure that the paper adopts for high-dimensional mediation testing.","marker":"Dai et al. (2022)"},{"why":"Provides the sparse and irregular longitudinal data model $y_{ij} = Y_i(t_{ij}) + \\varepsilon_{ij}$ that frames the outcome.","marker":"Zeng et al. (2021)"},{"why":"The function-on-scalar regression comparison method that finds no mediators in the case study.","marker":"Reiss et al. (2011)"},{"why":"Provides the polynomial spline estimation and correlation structure machinery for varying-coefficient longitudinal models.","marker":"Huang et al. (2004)"},{"why":"Supplies the weighted least squares weight matrix construction used in the spline fit.","marker":"Chu et al. (2016)"}],"fun_headline_variants":["Five lipids mediate sleep-cognition decline in older men","Sparse longitudinal mediation finds 5 lipid mediators for cognition","New mediation test for sparse longitudinal data pinpoints lipids","Sleep rhythm to cognitive decline: 5 lipid mediators identified","Mediation analysis for sparse longitudinal outcomes links lipids to cognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The FDR guarantee rests on the assumption that, for each mediator, the p-value for the exposure-to-mediator effect and the p-value for the time-varying mediator-to-outcome effect are asymptotically independent; the paper states this in Theorem 1 but gives no proof or sensitivity analysis for it.","fun_headline_variants_meta":{"raw":{"variants":["Five lipids mediate sleep-cognition decline in older men","Sparse longitudinal mediation finds 5 lipid mediators for cognition","New mediation test for sparse longitudinal data pinpoints lipids","Sleep rhythm to cognitive decline: 5 lipid mediators identified","Mediation analysis for sparse longitudinal outcomes links lipids to cognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000782,"raw_usage":{"total_tokens":3525,"prompt_tokens":1085,"completion_tokens":2440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2361}},"tokens_in":701,"tokens_out":2440,"duration_ms":22934,"temperature":1.0,"reasoning_tokens":2361,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:21:20.008419+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate null data with no true mediation but with shared subject-level random effects that couple the mediator and outcome equations, so $P_{\\alpha_k}$ and $P_{\\beta_k(t)}$ are dependent, and run the proposed permutation and mixture-FDR procedure with growing $p$; if the empirical FDR exceeds the target level as $p$ grows, the independence assumption behind Theorem 1 is violated.","supporting_citations":[],"review_version":1}