{"id":"88041252-1a2e-458e-8bb0-af6a4d56528f","arxiv_id":"2506.18415","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"Under transportability of the placebo cause-specific hazard, the authors derive efficient triply robust estimators for cumulative incidences with competing risks and show precision gains in simulations and in two cardiovascular trials.","lead":"This paper develops statistical methods that let clinical trials borrow external control data to estimate how treatments affect disease outcomes when patients can die before the health event happens, with formulas for the best possible precision gain. Applied to two cardiovascular trials, it narrows confidence intervals for the effect of a GLP-1 drug on non-fatal heart attack or stroke by about nine percent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's consistency condition (ii) omits a correct censoring hazard; with the stated condition the estimator is not consistent, so the triple-robustness claim is overstated.","rationale":"The reader's weak spot was the real-data transportability assumption and the unverifiable rate conditions in Assumption S2. I agree that those are genuine concerns for the applied illustration. However, the more load-bearing issue for the paper's central methodological claim is internal: Theorem 1 states a triple-robustness property, and one of its three consistency conditions, condition (ii), is not sufficient. The proof's own decomposition in Lemma S3 shows a non-vanishing first-order bias when the censoring hazard is misspecified, even though condition (ii) allows exactly that. This is a concrete, checkable defect in the stated theorem rather than a disagreement about external validity. It does not undermine the semiparametric efficiency-bound derivation, which appears sound, nor the estimator's behavior when all nuisance models are correct. But the paper's advertised robustness property is false as written, and the theorem needs a corrected condition and a revised proof before the methodological claim can be accepted. I would therefore keep the verdict conditional, but with the condition now including a mandatory correction to Theorem 1, not only the application-level caveats.","tokens_in":48466,"tokens_out":19832,"duration_ms":200992,"concrete_test":"Set up the minimal one-time-point model: no covariates, π=0.5, e₁(0|X)=1, true A•₁=0.1, A₁₂=0.2, S_c1=0.9, external A₀₂=0.3. Use the plug-in estimator with •₁=A•₁, π̂=π, ê₁=e₁, but deliberately misspecified Ā₁₂=0.4 and S̄_c1=1 (no censoring). Compute the probability limit P{ℓ̄₁(0)} − θ₁(0) either by evaluating the Lemma S3 remainder symbolically or by simulating the estimator with these fixed nuisance limits. If the bias is non-zero, condition (ii) of Theorem 1 is refuted; the same calculation with S̄_c1=S_c1 (but still wrong A₁₂) should give bias zero, confirming the missing condition.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Theorem 1, consistency is claimed under condition (ii): Ä•₁=A•₁, ē₁=e₁, π̄=π. This condition does not require the censoring hazard A_c1 or the cause-2 hazard A₁₂ to be correct. But in the proof, the bias remaining after (ii) is the second term of Lemma S3: P[π/α ∫ (b₂/b̄₂ − S•₁/S̄•₁) q̄₂ d(1 − S₁₂/S̄₁₂)]. Under (ii), S•₁/S̄•₁=1 and, since b₂ = e₁ S•₁ S_c1, we have b₂/b̄₂ = S_c1/S̄_c1. The term becomes P[π/α ∫ (S_c1/S̄_c1 − 1) q̄₂ d(1 − S₁₂/S̄₁₂)], which is generally non-zero when S_c1 and A₁₂ are misspecified. Neither is constrained by condition (ii), so consistency does not follow; for example, a misspecified Cox model for censoring together with a wrong cause-2 hazard leaves a first-order bias. The statement should include ̄S_c1=S_c1 in (ii); with that addition the remainder vanishes independently of A₁₂. The efficiency-bound result and the full-consistency conclusion are unaffected, but the advertised triple robustness is not established as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops semiparametric theory for estimating causal cumulative incidences and treatment effects in a randomized trial augmented with external controls, in a competing-risks setting. The target parameters are the causal cumulative incidences θ_j(a) = pr{T(a) ≤ τ, J(a) = j | D = 1} and their differences, plus restricted mean times lost. The key structural assumption is transportability of the conditional cause-1 hazard under control between the trial population and the external population (Assumption 4). Under this assumption, the authors derive efficient influence functions for θ_1(0) and θ_2(0) (Lemma 1, with the general form in Proposition S1), quantify the theoretical efficiency gain from fusion (Corollary 1), and construct plug-in EIF estimators. Theorem 1 claims triple robustness for θ̂_1(0) and asymptotic linearity at the efficiency bound when all nuisance models are consistent and certain rate conditions hold. The paper includes a simulation study and an application to SUSTAIN-6 and LEADER cardiovascular outcome trials.","tokens_in":48725,"tokens_out":8614,"duration_ms":85943,"significance":"If the robustness statement is corrected, this is a valuable contribution: it extends data-fusion efficiency theory from continuous or binary outcomes to time-to-event outcomes with competing risks, under a transportability assumption that is weaker than full distributional exchangeability and has a clear causal interpretation. The supplementary proofs are detailed and self-contained, the efficiency-bound expression in Corollary 1 is derived rather than asserted, and the simulations show low bias, good agreement between model-based standard errors and empirical RMSE, and near-nominal coverage in the correctly specified setting. The real-data analysis is thoughtful and includes a sensitivity analysis that openly flags the fragility of the transportability assumption, which is a strength of the paper.","major_comments":[{"comment":"The consistency condition (ii) is insufficient as stated. Condition (ii) requires only Ā•1=A•1, ē1=e1, and π̄=π, not a correct censoring hazard A_c1. Substituting condition (ii) into the remainder term of Lemma S3 (Supplementary S2.3) leaves the second term P[(π/α)(S_c1/S̄_c1 − 1) q̄2 d(1 − S12/S̄12)], which is generally non-zero when S̄_c1 and S̄12 are misspecified. Since neither the censoring hazard nor the cause-2 hazard is constrained by condition (ii), consistency does not follow; for example, a misspecified Cox model for censoring combined with a wrong cause-2 hazard produces a first-order bias. The condition should include S̄_c1 = S_c1 (equivalently Ā_c1 = A_c1); with that addition the remainder vanishes independently of S̄12. Consequently, the sentence after Theorem 1 stating that consistency requires only correct estimation of at least one of the cause-specific hazards is overstated, and the triple-robustness property is not established as written. The efficiency-bound result and the fully consistent asymptotic-linearity result are unaffected, but the advertised robustness set must be corrected.","section":"Section 3.3, Theorem 1"}],"minor_comments":[{"comment":"In the sentence defining the model P, 'the distribution of A ≡ 0 is degenerate when D = 1' appears to be a typo: the external population D = 0 contributes only controls, whereas the target population D = 1 is an RCT with both A = 0 and A = 1. The sentence should read 'when D = 0'.","section":"Section 3.2"},{"comment":"The explanatory paragraph following Theorem 1 should be revised after correcting condition (ii). As written, it suggests that correct estimation of A•1 alone suffices for consistency, but the corrected condition requires the control-arm censoring hazard A_c1 to be correct as well.","section":"Section 3.3, after Theorem 1"},{"comment":"The simulation study fits all nuisance models correctly, so it does not exercise the misspecification branches of Theorem 1. A scenario with a misspecified censoring hazard, possibly together with a misspecified cause-2 hazard, would directly illustrate the corrected triple-robustness property and clarify the practical consequences of the condition (ii) fix.","section":"Section 4"},{"comment":"For n = 750, the coverage for θ_1(0,t) at t = 1 is 92.3–92.7%, slightly below the nominal 95%; a brief comment on this mild undercoverage would be helpful for readers who focus on small-sample operating characteristics.","section":"Table 1"},{"comment":"The tie-breaking jitter scale is set to 10^{-5} without a sensitivity analysis. Since Assumption 5 is an asymptotic condition, a short sensitivity check varying the jitter scale would strengthen the real-data analysis and reassure readers that the reported precision gains are not artifacts of the tie-breaking choice.","section":"Supplementary S4"}],"recommendation":"major_revision","confidential_remarks":"The triple-robustness gap identified in the stress-test note is real and load-bearing, but it is local and fixable: adding S̄_c1 = S_c1 to condition (ii) restores consistency of the stated branch. I would not reject on this basis. The rest of the theoretical development appears sound, and the paper already contains the machinery needed to make the correction. I would encourage the authors to also add a simulation with a misspecified censoring hazard and to soften the abstract's claim about the real-data standard error reduction, since that reduction is conditional on an unverifiable transportability assumption that the paper's own sensitivity analysis shows to be fragile."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, read the paper. Main result: first semiparametric efficiency bound for cumulative incidence with external controls under hazard transportability, with a plug-in EIF estimator and a variance-reduction formula. The theory is mostly solid, but there's a hole in Theorem 1's triple-robustness claim that needs fixing.\n\nThe genuinely new pieces are Lemma 1's EIF for theta_1(0), Corollary 1's variance reduction expression, and the construction combining trial and external controls through H-dot weighted martingale integrals. The supplement has detailed proofs, and the simulation study is careful: low bias, RMSE-SE agreement, coverage near nominal. The real-data analysis is honest, including a sensitivity analysis that omits cardiovascular history and shows point estimates move—good to see that flagged.\n\nThe problem: Theorem 1 states consistency under condition (ii): correct cause-1 hazard, treatment propensity, and selection score, but says nothing about the censoring hazard or cause-2 hazard. The proof claims the bias vanishes by Lemma S3. Under (ii), the first remainder term does vanish, but the second reduces to P[pi/alpha * integral of (S_c1/S_bar_c1 - 1) q_bar_2 d(1 - S_12/S_bar_12)]. Unless the censoring hazard is also correct, this term is generally non-zero; a wrong cause-2 hazard leaves a first-order bias. So the estimator is not consistent under the stated condition. The fix is simple: add S_bar_c1 = S_c1 to condition (ii). With that, the remainder vanishes regardless of A_12. The efficiency bound and the fully-consistent asymptotic-linearity result are unaffected, but the advertised triple robustness is not established as written. This is a theorem-level error, not a cosmetic one.\n\nOther soft spots: the transportability assumption is strong and untestable; the real-data gain depends on pooling two different placebos and on the covariate set being sufficient. The paper's own sensitivity analysis shows that omitting CVD history shifts point estimates materially, so the 9% CI shortening is conditional on an assumption that may not hold. Asymptotic linearity also needs rate conditions (Assumption S2) that the simulations don't stress, since they use correctly specified Cox models. No code or data is provided.\n\nWho should read this: anyone working on hybrid control designs or data fusion for time-to-event outcomes. It deserves peer review, but the triple-robustness statement needs correction before I'd trust the theorem as stated. I'd send it to a statistician who knows semiparametric theory.","headline":"The efficiency-bound result for external-control fusion in competing risks is a real contribution, but Theorem 1's triple-robustness claim has a gap: condition (ii) as written does not imply consistency unless the censoring hazard is also correct.","tokens_in":49295,"tokens_out":4252,"would_cite":true,"duration_ms":41315,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","62N02"],"pacs":[],"model":"deepseek-v4-flash","headline":"Augmenting a randomized trial with external controls provably lowers the best achievable variance of cumulative-incidence estimates, and a new triply robust estimator attains that bound.","keywords":["competing risks","external controls","data fusion","semiparametric efficiency bound","transportability","cumulative incidence","triply robust estimation","restricted mean time lost"],"falsifier":"Take the SUSTAIN-6/LEADER data and fit the fused and trial-only estimators with and without cardiovascular history in the covariate set (the paper's Tables S8-S9 do exactly this): the week-104 fused estimate moves from -2.72 to -2.44 percentage points while the trial-only estimate moves from -2.56 to -2.88, so the two estimates swap order and the gap exceeds the 9-percent CI narrowing. Any single covariate whose addition or removal flips the sign of the fused-minus-trial-only difference by more than the confidence-interval shortening, or a Cox model of the placebo cause-1 hazard with significant trial-by-covariate interactions, falsifies the transportability assumption and shows the gain is precision bought with bias.","tokens_in":48246,"feed_emoji":"📊","tokens_out":17513,"duration_ms":143374,"temperature":0.7,"pith_summary":"This paper shows that augmenting a randomized trial's control arm with external controls can lower the variance of causal cumulative-incidence estimates under competing risks, provided one assumption holds: after conditioning on baseline covariates, the hazard of the event of interest under control is identical in the trial and the external population. Under this transportability assumption, the authors derive the semiparametric efficiency bound for the cumulative incidences $\\theta_j(a)$ — the variance floor any regular estimator must respect — and express the variance reduction from borrowing external information as a closed-form integral (Corollary 1). They construct a plug-in estimator from the efficient influence function that is triply robust (consistent when any of three sets of nuisance models is correctly specified) and asymptotically efficient when all nuisance models are (Theorem 1). In the SUSTAIN-6/LEADER application, fusing the external placebo arm shortens confidence intervals for the effect of semaglutide on non-fatal cardiovascular events by about 9 percent, and by more than 20 percent for the control-arm cumulative incidence itself.","feed_headline":"9% shorter CIs from fusing external trial controls — bound proven","feed_subtitle":"Derives the max precision gain from external control arms — and the estimator that achieves it.","key_machinery":"The engine of the argument is the efficient influence function $\\varphi_1(0)$ for the control-arm cumulative incidence in the transportability model. Its event-of-interest martingale integral is weighted by the inverse of the pooled at-risk probability $H_\\bullet(t|x)=\\pi(x)e_1(0|x)(S_1S_1^c)(t|0,x)+\\{1-\\pi(x)\\}(S_0S_0^c)(t|x)$, a mixture of trial-control and external-control survival-and-censoring products weighted by the selection score $\\pi(x)$; this pooled weight is precisely how the external sample enters the estimator, while the cause-2 hazard and cumulative-incidence terms remain population-specific. The transportability assumption itself is the bridge: it equates the cause-1 hazard under control across populations after conditioning on $X$, so the external sample can be used to estimate the shared hazard $A_{\\bullet 1}(t|0,x)$, and Corollary 1's reduction formula quantifies the gain as an integral of a positive term against $dA_{11}(t|0,x)$, which is larger when the competing-risk or censoring rate is higher in the trial control arm than in the external population. Theorem 1's triple robustness follows from the von Mises expansion of the plug-in estimator, whose remainder terms vanish under any of three correct-model combinations and at $o_P(n^{-1/2})$ rates under the full correct specification (Assumption S2).","core_discovery":"The paper's central claim is that, under transportability of the conditional cause-1 hazard under control (Assumption 4, $A_{11}(0)(t|x)=A_{01}(t|x)$), the semiparametric efficiency bound for the causal cumulative incidence $\\theta_1(0)$ is characterized by the variance of an explicit efficient influence function (Lemma 1), and that this bound is strictly smaller than the bound without external controls precisely when covariate overlap between the populations is positive and subjects remain at risk of the event over a non-trivial time span (Corollary 1). The proposed estimator $\\hat{\\theta}_1(0)$ is the empirical mean of the plug-in efficient influence function; it is consistent whenever one of three groups of nuisance models is correctly specified — the cause-1 hazard, the cause-2 hazards, or the selection score together with the censoring hazards — and it is asymptotically linear and attains the bound when all nuisance models are consistent (Theorem 1). The real-data analysis applies this machinery to two cardiovascular outcome trials: using LEADER's placebo arm as external control for SUSTAIN-6, the confidence interval for the effect on non-fatal cardiovascular events with all-cause death as competing risk narrows by about 9 percent, with point estimates essentially unchanged, while the under-placebo cumulative incidence gains more than 20 percent.","pith_inferences":["Read as a design-stage decision rule, the Corollary 1 formula can tell a trial planner when NOT to bother fusing: if covariate overlap is thin or the competing-risk/censoring imbalance is small, the theoretical gain shrinks toward zero, so the formula doubles as a cost-benefit tool — a use the paper leaves implicit.","The 9 percent real-data gain is specific to a trial with a large control arm; the paper's downsampled analysis shows 45-50 percent CI shortening, which suggests the method's practical payoff is concentrated in rare-disease and small-trial settings where control observations are scarce.","A concrete pre-registration rule follows: keep the fused estimate only if the fused and trial-only point estimates agree within the expected CI narrowing; the paper proposes a formal test-then-pool estimator as future work, and this agreement check is the operational version.","The transportability assumption is asymmetric — it constrains only the event-of-interest hazard, leaving the competing hazard and censoring free to differ — and that asymmetry is likely what makes the assumption defensible in real trials; it also suggests the efficiency-bound machinery could transfer to other multi-state estimands such as average hazards with survival weights, which the paper poin"],"forward_implications":["Because Corollary 1 expresses the variance reduction as a closed-form integral, a trial planner can compute the theoretical ceiling on the precision gain from external controls before collecting data, given assumed hazards, censoring rates, and covariate overlap.","The fusion estimator is triply robust: consistent estimation of the control-arm cumulative incidence survives misspecification of the selection score as long as the cause-1 hazard is correct, and vice versa, so the method does not require every nuisance model to be right.","The efficiency gain for treatment effects is bounded by the active-arm sample size, not the control-arm size: with external controls nearly three times the trial controls, confidence intervals still shorten by only about 9 percent, while under-placebo parameters gain over 20 percent.","When all nuisance models are consistent and the unverifiable rate conditions in Assumption S2 hold, the estimator is asymptotically linear and attains the semiparametric efficiency bound.","By the chain rule, the same efficient-influence-function machinery extends to restricted mean times lost (Corollary 2), so the precision gains transfer to that estimand."],"supporting_citations":[{"why":"Supplies the efficient influence function for the active-arm and trial-only estimators that the fusion estimator generalizes and is benchmarked against.","marker":"Rytgaard et al. (2023)"},{"why":"The SUSTAIN-6 cardiovascular outcome trial data, which define the target population in the real-data analysis.","marker":"Marso et al. (2016a)"},{"why":"The LEADER trial data, which provide the external placebo controls fused into the target trial.","marker":"Marso et al. (2016b)"},{"why":"Sets out the non-nested sampling design used to frame the asymptotic model for combining trial and external samples.","marker":"Dahabreh et al. (2021)"},{"why":"Provides the counting-process martingale and product-integral identities used in deriving the efficient influence function and its asymptotics.","marker":"Fleming and Harrington (1991)"},{"why":"Supplies the Duhamel and backward product-integral equations used to expand the plug-in estimator in the proof of Theorem 1.","marker":"Gill and Johansen (1990)"},{"why":"Provides the empirical-process lemma establishing asymptotic linearity of the plug-in estimator.","marker":"van der Vaart (1998)"},{"why":"The subdistribution-hazard approach rejected as a transportability assumption, motivating the paper's choice of cause-specific hazard transportability.","marker":"Fine and Gray (1999)"}],"fun_headline_variants":["Fusing external controls shrinks trial CIs by 9%","Semiparametric bound proves external-control gains","Triply robust estimator achieves efficiency bound","External control arms: precision bound achieved","9% tighter CIs via transportable hazards"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole precision gain rests on the assumption that, after adjusting for the measured baseline covariates, the placebo rate of the event of interest is genuinely the same in the external population as in the trial — and, in the real-data example, that a once-daily and a once-weekly placebo injection can be treated as the same control.","fun_headline_variants_meta":{"raw":{"variants":["Fusing external controls shrinks trial CIs by 9%","Semiparametric bound proves external-control gains","Triply robust estimator achieves efficiency bound","External control arms: precision bound achieved","9% tighter CIs via transportable hazards"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000338,"raw_usage":{"total_tokens":1916,"prompt_tokens":1038,"completion_tokens":878,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":807}},"tokens_in":654,"tokens_out":878,"duration_ms":8126,"temperature":1.0,"reasoning_tokens":807,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:49:42.670125+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the SUSTAIN-6/LEADER data and fit the fused and trial-only estimators with and without cardiovascular history in the covariate set (the paper's Tables S8-S9 do exactly this): the week-104 fused estimate moves from -2.72 to -2.44 percentage points while the trial-only estimate moves from -2.56 to -2.88, so the two estimates swap order and the gap exceeds the 9-percent CI narrowing. Any single covariate whose addition or removal flips the sign of the fused-minus-trial-only difference by more than the confidence-interval shortening, or a Cox model of the placebo cause-1 hazard with significant trial-by-covariate interactions, falsifies the transportability assumption and shows the gain is precision bought with bias.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the efficient influence function for the active-arm and trial-only estimators that the fusion estimator generalizes and is benchmarked against."},{"cited_title":"J., Haneuse, S","cited_arxiv_id":null,"evidence_quote":"Sets out the non-nested sampling design used to frame the asymptotic model for combining trial and external samples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the counting-process martingale and product-integral identities used in deriving the efficient influence function and its asymptotics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the empirical-process lemma establishing asymptotic linearity of the plug-in estimator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The subdistribution-hazard approach rejected as a transportability assumption, motivating the paper's choice of cause-specific hazard transportability."}],"review_version":2}