{"id":"6b237c60-3b47-4df5-9b70-40521f248aed","arxiv_id":"2411.18739","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A BART-based G-computation framework estimates interventional direct and indirect effects for longitudinal mediators and survival outcomes, extends to competing events, and finds null short-term BP medication effects in ARIC.","lead":"This paper develops a Bayesian BART-based method for estimating how much of a treatment's effect on time-to-death runs through a time-varying risk factor, with time-varying confounders and competing events handled. It applies the method to blood-pressure medication and CVD death in the ARIC cohort, finding no significant short-term effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim relies on Assumptions 3(ii)/6(ii), mediator-outcome sequential ignorability, which the authors themselves flag as suspect; the Section 8 CMBP correction addresses only one confounder and omits its own estimation/transportability uncertainty, so this vulnerability remains…","rationale":"The reader correctly identifies Assumption 3(ii)/6(ii), no unmeasured confounding of the mediator-outcome relationship, as the weakest load-bearing assumption, and the authors' own Sections 8 and 9 corroborate this. My reading adds specificity: the proposed external-data correction for cumulative mean blood pressure does not constitute a sensitivity analysis because it fixes the CARDIA model coefficients and randomly sampled random effects, ignores transportability uncertainty, and addresses only one of several named confounders. The identification proofs and G-computation algorithms are internally coherent and the simulation checking of the estimation procedure under correct models is genuine supporting evidence, so I do not see an internal inconsistency or a fatal technical flaw. The conditionality of the reader's verdict is appropriate: the method is adoptable once the unmeasured-confounding limitation is either addressed quantitatively or stated more prominently as a scope restriction. My stress-test read therefore leaves the verdict unchanged.","tokens_in":76734,"tokens_out":4590,"duration_ms":50520,"concrete_test":"Conduct a calibrated simulation with a known data-generating process in which an unmeasured covariate U affects both the mediator M_t and the survival outcome while being omitted from the fitted BART models. Vary the strength of U (e.g., effect on M_t of 0 to 1 SD per unit U, and hazard ratio for CVD death of 1 to 3 per SD of U) and its prevalence, then estimate IIE/IDE with the proposed algorithm. Report bias and coverage versus the true values as a function of confounding strength. If modest U effects produce material bias or coverage failures in the direction needed to explain the ARIC estimates, then Assumption 3(ii) is demonstrably load-bearing and the paper should report a quantitative sensitivity/bounding analysis rather than the current CMBP check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Propositions 1a and 2 claim identification of interventional direct, indirect, and total effects on survival and on cause-specific CIFs from observed data. In the ARIC application the decisive identifying condition is Assumption 3(ii)/6(ii): conditional on observed history, the mediator (mean blood pressure) is independent of all subsequent covariates and outcomes. The authors explicitly identify plausible violations (Section 8: pre-baseline hypertension history; Section 9: diet, exercise, medication adherence, family history). The Section 8 remedy does not close this gap. First, CMBP is only one of several suspected confounders. Second, Assumption 7, transportability of the CARDIA mixed-effects model to ARIC conditional on race and sex, is asserted and untestable. Third, the imputation procedure samples one random-effects pair per ARIC participant from race/sex matches but does not propagate uncertainty in the fitted fixed effects, the random-effects distribution, or the matching step; the paper states this explicitly and does not pursue it. Fourth, the comparison of Tables 3-4 with Tables 5-6 is therefore not a quantitative sensitivity analysis: an unchanged point estimate under one, uncertainty-free imputation does not bound the bias from an unmeasured confounder. Because the mediation contrasts are not identified when Assumption 3(ii) fails, the intended causal interpretation of the reported IIE/IDE is not secured. This is a limitation rather than an internal inconsistency: the identification proofs appear to follow given the assumptions, the G-computation algorithms are coherent with those proofs, and the simulations validate estimation under correct models, but they do not address failure of the no-unmeasured-confounding assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a Bayesian semi-parametric approach to causal mediation for longitudinal mediators and time-to-event outcomes, with both a no-competing-event and a competing-event version. The estimands are interventional direct and indirect effects on survival probabilities and on cause-specific cumulative incidence functions, defined via random interventions on the mediator. Identification is claimed under standard sequential randomization, positivity, and consistency assumptions (Propositions 1a and 2). Estimation uses BART for the observed-data models and a G-computation algorithm that propagates posterior samples through the sequential conditional distributions. The method is applied to the ARIC cohort to estimate the effect of blood pressure medication on time to CVD death, mediated by mean blood pressure, with smoking as a time-varying confounder and non-CVD death as a competing event. A separate section proposes imputing cumulative mean blood pressure from the external CARDIA cohort to reduce suspected unmeasured confounding.","tokens_in":77063,"tokens_out":5525,"duration_ms":56961,"significance":"If the results hold, the paper offers a useful and flexible implementation of interventional mediation analysis for a setting that is difficult to handle with parametric structural equation models: time-varying exposures, mediators, and confounders, with discrete-time survival and competing risks. The BART-based G-computation is a natural extension of existing work by Zheng and van der Laan (2017), Lin et al. (2017), and Vansteelandt et al. (2019), and the authors provide R code, which is a practical asset. The identification proofs follow standard sequential randomization arguments and are largely correct in structure. The main weaknesses are that the simulation studies validate the estimator against a self-generated GLM truth rather than against a known data-generating mechanism, and that the proposed external-data correction for unmeasured confounding does not quantitatively bound the bias from violations of the mediator-outcome ignorability assumptions, which the authors themselves acknowledge in Section 9.","major_comments":[{"comment":"The CMBP correction does not close the gap left by Assumption 3(ii)/6(ii). Section 8 imputes only one suspected unmeasured confounder, while Section 9 lists diet, physical activity, family history, and medication duration/adherence as additional plausible confounders. Moreover, the comparison of Tables 3-4 with Tables 5-6 is not a quantitative sensitivity analysis: the imputation procedure samples one random-effects pair per ARIC participant from race/sex matches but does not propagate uncertainty in the fitted fixed effects, the random-effects distribution, or the matching step, and the authors state this explicitly. An unchanged point estimate under such an imputation does not bound the bias from an unmeasured confounder, so the conclusion that \"the results are not very sensitive\" in Section 8.2 overstates what is established. The estimands remain identified only under the untestable sequential ignorability plus transportability assumptions; the Section 8 analysis should be presented as exploratory rather than as a sensitivity bound.","section":"Section 8.1/8.2 and Assumption 7"},{"comment":"Step 2(b)(vi) instructs the user to \"sample one observation of Y_{t_{j+1}},(1)\" from a Bernoulli distribution with probability I^(2), the hazard for the competing event. This is internally inconsistent: the main-event indicator should be sampled from I^(1), and the competing-event indicator should be sampled from I^(2). As written, the sequential G-computation for competing events would draw the wrong event indicators and would not implement the likelihood factorization in equation (10). If this is a typographical error, the step must be corrected and the code checked against the corrected algorithm; if not, the competing-events implementation is incorrect.","section":"Algorithm 2, Step 2(b)(vi)"},{"comment":"The simulation ground truth is generated from GLMs fitted to the observed ARIC data, so the \"true\" causal effects are functionals of parametric fits to the same sample, not quantities from a known structural data-generating process. The misspecification scenario adds log and interaction terms but remains within the same generalized linear framework and does not alter the confounding assumptions. Consequently, the simulations demonstrate that BART can approximate the target of a GLM-based G-computation and is more robust to certain nonlinearities, but they do not evaluate bias when Assumption 3(ii)/6(ii) fails, which is the principal threat in the application. I recommend adding a simulation with a known data-generating mechanism, or with a controlled unmeasured mediator-outcome confounder, or explicitly delimiting what the current simulations establish.","section":"Section 6, Tables 1-2"},{"comment":"Proposition 2 states identification under \"Assumptions 1-6,\" but Assumptions 4-6 are introduced as modifications of Assumptions 1-3 for competing events. It is not clear whether Assumptions 1-3 are also required in the competing-event setting or are entirely replaced by Assumptions 4-6. The proof in the supplement appears to use consistency and no-unmeasured-confounding conditions in the competing-event form, but the statement should be clarified so that readers know exactly which set of assumptions is being invoked.","section":"Section 4, Proposition 2"}],"minor_comments":[{"comment":"The text refers to \"the model in (24)\" when describing the CARDIA mixed-effects model, but the model is displayed as equation (18).","section":"Section 8.1"},{"comment":"Algorithm 1 refers to models \"as specified in Section 4.2,\" but the model specification is in Section 5.2.","section":"Algorithm 1, Step 1"},{"comment":"The table note contains a stray closing parenthesis: \"in ARIC. )\" should read \"in ARIC.\"","section":"Table 10 note"},{"comment":"The simulation section does not report the number of Monte Carlo repetitions used for the G-computation step or the number of posterior draws kept for the BART and parametric models, which would be helpful for assessing Monte Carlo error.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"For the editor: The identification framework closely follows Zheng and van der Laan (2017) and the recent TMLE work by Wang et al. (2023); the main novel contributions are the BART-based implementation, the competing-event extension, and the CARDIA-based CMBP imputation. The CMBP imputation is currently more of an exploratory check than a quantitative sensitivity analysis, and Algorithm 2 contains a potentially consequential typo. The manuscript is within scope for a statistical methodology journal, but the application is underpowered (172 CVD deaths) and the discussion of the application should be tempered accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something useful: it puts interventional mediation estimands for time-varying exposures, longitudinal mediators, and survival outcomes into a BART-based G-computation framework, and adds a competing-risks version. The estimands themselves aren't new—Zheng and van der Laan, Lin et al., and Vansteelandt et al. have the core ideas—but the specific combination with BART and competing risks is, and the implementation is careful. The identification proofs are standard sequential randomization arguments, and they check out as far as I can see. The G-computation algorithms are explicit and the code is provided. The simulation study is internally consistent: BART does well relative to a parametric model when the parametric model is misspecified, and both are roughly unbiased when it isn't. That's genuine evidence for the estimator, though not for the causal question.\n\nThe soft spots. First, the simulation ground truth is generated from GLMs fit to the same ARIC data, so it's really a test of BART versus a parametric model on self-generated data, not a validation against external truth. That's a limitation, not a fatal one, for a methods paper. Second, the ARIC application requires sequential ignorability for the mediator-outcome relationship (Assumptions 3(ii)/6(ii)), and the authors themselves say this is suspect. Diet, exercise, adherence, family history—plausible confounders. The CARDIA-based CMBP adjustment in Section 8 addresses one confounder, assumes transportability without being able to test it, and doesn't propagate imputation uncertainty. The authors are transparent about this, but the \"results are not sensitive\" claim is weaker than it looks: an unchanged point estimate under one uncertainty-free imputation doesn't bound the bias from an unmeasured confounder. The causal interpretation of the reported IIE/IDE is therefore not secured for this application. That should be stated more sharply in the paper, not just in the discussion.\n\nWho is this for? Applied statisticians and epidemiologists working on mediation with longitudinal mediators and survival outcomes, especially those interested in BART flexibility and competing risks. It's a solid within-subfield contribution, not a breakthrough. The central methodology holds up; the application's causal claims need caveats.\n\nRecommendation: send it to a serious referee. It deserves the time. The authors should be asked to either strengthen the sensitivity analysis or clearly label it as an illustration, and to discuss the simulation's self-generated truth more humbly.","headline":"A coherent BART-based implementation of interventional mediation for longitudinal and survival data with a competing-risks extension; the application's causal interpretation rests on assumptions the authors themselves flag, but the methodology is solid and worth a serious referee.","tokens_in":77641,"tokens_out":2291,"would_cite":true,"duration_ms":23893,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62F15","62N01"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian semi-parametric G-computation approach identifies interventional direct and indirect effects on time-to-event outcomes with time-varying exposures, mediators, and confounders, and extends the identification to competing risks.","keywords":["causal mediation","interventional direct and indirect effects","Bayesian additive regression trees","time-to-event outcomes","competing risks","time-varying confounding","longitudinal data","G-computation"],"falsifier":"Obtain a measured proxy for the suspected confounder, such as self-reported duration of hypertension or medication adherence, and re-estimate the model with it added to the mediator and outcome BART models; if either the estimated interventional direct or indirect effect shifts by more than one credible-interval width, the no-unmeasured-confounding assumption is contradicted by the data.","tokens_in":76522,"feed_emoji":"💊","tokens_out":9683,"duration_ms":82932,"temperature":0.7,"pith_summary":"This paper develops causal mediation analysis for longitudinal cohort data in which exposure, mediator, and confounders are measured repeatedly and the outcome is a time-to-event such as death. The authors establish that interventional direct, indirect, and total effects on survival—and, with competing events, on cause-specific cumulative incidence functions—are identified from observed data under sequential no-unmeasured-confounding assumptions. They then estimate these effects with a semi-parametric Bayesian additive regression tree (BART) G-computation algorithm that accommodates continuous or binary mediators and confounders without linearity or proportional-hazards assumptions. Applied to blood-pressure medication and cardiovascular death in a hypertensive cohort, the method yields effect estimates that are mostly statistically indistinguishable from zero, which the authors attribute to short follow-up and possible violation of strong ignorability.","feed_headline":"Flexible estimator separates direct and indirect survival effects","feed_subtitle":"A Bayesian tree-based G-computation approach handles longitudinal mediators, confounders, and competing risks.","key_machinery":"The load-bearing object is the interventional mediator distribution: for a subject with a given observed longitudinal history, the mediator is not fixed to its counterfactual level but drawn randomly from the distribution of mediator values among subjects under the control exposure regime with matching history. This sidesteps the non-identification of natural direct and indirect effects when time-varying confounders are affected by prior exposure. Estimation is carried out by BART models for each time-varying confounder, mediator, and discrete-time hazard, with a probit link for binary variables, together with a Bayesian bootstrap over baseline confounders. For competing events, the joint event likelihood is decomposed into a BART model for the main event hazard and a second BART model for the competing-event hazard conditional on no main event, avoiding proportional-hazards assumptions.","core_discovery":"The central claim is that interventional direct and indirect effects are identified for time-varying exposures, mediators, and confounders with a survival outcome. For exposure regimes $z$ and $z^*$, the estimator contrasts survival probabilities under a static exposure intervention and a random mediator intervention: the interventional direct effect is $\\mathrm{IDE}(t_j) = S_{z,z^*}(t_j) - S_{z^*,z^*}(t_j)$, the indirect effect is $\\mathrm{IIE}(t_j) = S_{z,z}(t_j) - S_{z,z^*}(t_j)$, and they sum to the total effect. The paper proves that these contrasts reduce to functionals of the observed data distribution, and the competing-event extension identifies the analogous contrasts for cause-specific cumulative incidence functions for both the main and the competing event. A BART-based G-computation algorithm then estimates the required conditional densities and integrates them over the observed baseline confounder distribution.","pith_inferences":["Editorial extension: because identification rests on sequential no-unmeasured-confounding, a formal sensitivity analysis that perturbs the mediator-outcome confounding structure would quantify how far IDE and IIE can move before the causal reading fails.","Editorial extension: the two-model competing-event decomposition extends naturally to more than two event types by adding further conditional BART hazard models, and the same G-computation steps would apply.","Editorial extension: the external-cohort imputation of cumulative mean blood pressure assumes transportability of a blood-pressure trajectory model across cohorts; that assumption could be tested by validating predicted trajectories against later observed blood pressure in the target cohort.","Editorial extension: incorporating predicted mediator values into the outcome model, the route the authors note as future work for reducing BART 'prior dogmatism,' could adapt this estimator to higher-dimensional covariate settings."],"forward_implications":["If the identifying assumptions hold, the estimator gives approximately unbiased interventional direct, indirect, and total effects for time-to-event outcomes without assuming linearity, proportional hazards, or a correctly specified parametric outcome model.","The competing-event version provides a decomposition of total effects on cause-specific cumulative incidence functions, which is useful when non-cardiovascular death precludes the event of interest.","Simulation results indicate that, when the data-generating model includes interactions and log terms, the BART estimator has smaller bias and better credible-interval coverage than a misspecified parametric model.","In the cardiovascular application, the estimated direct and indirect effects of blood-pressure medication on death from cardiovascular disease are mostly null over the short follow-up window, consistent with the small number of events rather than with evidence of no mediation.","Adding an externally predicted cumulative blood-pressure history as a baseline confounder does not materially change the estimates under the stated transportability assumption, suggesting that this particular unmeasured confounder is not driving the conclusions."],"supporting_citations":[{"why":"Supplies the interventional-effects identification framework and sequential randomization assumptions that Proposition 1 builds on.","marker":"Zheng and van der Laan (2017)"},{"why":"Introduced interventional direct and indirect effects as random-intervention analogues that avoid recanting-witness non-identification.","marker":"VanderWeele and Tchetgen Tchetgen (2017)"},{"why":"A prior mediation approach for survival outcomes with time-varying exposures and mediators that this paper extends by conditioning mediator draws on longitudinal confounders.","marker":"Lin et al. (2017)"},{"why":"Defines BART, the ensemble-of-trees model used for time-varying confounders and mediators.","marker":"Chipman et al. (2010)"},{"why":"Provides the nonparametric discrete-time competing-risk BART framework that the survival outcome models in Section 5.2.3 extend.","marker":"Sparapani et al. (2020)"},{"why":"Established flexible Bayesian semi-parametric G-computation for longitudinal cohort data, which this paper adapts to mediation and survival.","marker":"Josefsson and Daniels (2021)"},{"why":"Gives the Bayesian bootstrap used in Algorithms 1 and 2 to sample baseline confounder vectors during G-computation.","marker":"Rubin (1981)"},{"why":"Clinical evidence that pre-baseline blood-pressure history confounds the medication-outcome relationship, motivating the cumulative-mean-blood-pressure adjustment.","marker":"Whelton et al. (2018)"}],"fun_headline_variants":["Bayesian trees separate survival direct and indirect effects","BART mediation for longitudinal data and survival endpoints","New estimator for causal mediation with competing risks","Separating direct and indirect survival effects with BART","Flexible survival mediation with time-varying confounders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, after conditioning on everything observed up to a visit, no unmeasured factor such as diet, exercise, medication adherence, family history, or pre-study hypertension history affects both the mediator and the outcome; if that fails, the estimated direct and indirect effects are not causal.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian trees separate survival direct and indirect effects","BART mediation for longitudinal data and survival endpoints","New estimator for causal mediation with competing risks","Separating direct and indirect survival effects with BART","Flexible survival mediation with time-varying confounders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1440,"prompt_tokens":909,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":525,"tokens_out":531,"duration_ms":111071,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:55:49.399006+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Obtain a measured proxy for the suspected confounder, such as self-reported duration of hypertension or medication adherence, and re-estimate the model with it added to the mediator and outcome BART models; if either the estimated interventional direct or indirect effect shifts by more than one credible-interval width, the no-unmeasured-confounding assumption is contradicted by the data.","supporting_citations":[{"cited_title":"and van der Laan, M","cited_arxiv_id":null,"evidence_quote":"Supplies the interventional-effects identification framework and sequential randomization assumptions that Proposition 1 builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced interventional direct and indirect effects as random-intervention analogues that avoid recanting-witness non-identification."},{"cited_title":"G., Logan, R., and VanderWeele, T","cited_arxiv_id":null,"evidence_quote":"A prior mediation approach for survival outcomes with time-varying exposures and mediators that this paper extends by conditioning mediator draws on longitudinal confounders."},{"cited_title":"A., George, E","cited_arxiv_id":null,"evidence_quote":"Defines BART, the ensemble-of-trees model used for time-varying confounders and mediators."},{"cited_title":"R., McCulloch, R","cited_arxiv_id":null,"evidence_quote":"Provides the nonparametric discrete-time competing-risk BART framework that the survival outcome models in Section 5.2.3 extend."},{"cited_title":"and Daniels, M","cited_arxiv_id":null,"evidence_quote":"Established flexible Bayesian semi-parametric G-computation for longitudinal cohort data, which this paper adapts to mediation and survival."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Bayesian bootstrap used in Algorithms 1 and 2 to sample baseline confounder vectors during G-computation."},{"cited_title":"K., Carey, R","cited_arxiv_id":null,"evidence_quote":"Clinical evidence that pre-baseline blood-pressure history confounds the medication-outcome relationship, motivating the cumulative-mean-blood-pressure adjustment."}],"review_version":1}