{"id":"bee05982-6a2f-48d5-a223-7c381775d77f","arxiv_id":"2505.15944","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"For estimating average treatment effects in a target population, a covariate-dependent allocation based on conditional outcome variance is provably optimal regardless of the target covariate distribution and the data configuration.","lead":"This paper shows how to choose the randomization ratio in a clinical trial when the goal is to estimate treatment effects in a patient population different from the trial itself. The authors find that a covariate-dependent allocation rule is optimal for any target population, so one design can serve both internal and external validity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The invariance result is correct under mean exchangeability, but that assumption is never stated formally and the abstract's 'regardless' phrasing overstates its robustness.","rationale":"The reader's verdict of ACCEPT is appropriate. The paper's mathematical results are internally coherent: under the usual exchangeability framework, minimizing the variance of the efficient influence function for Delta* gives popt(W) pointwise, and the r(W) factors cancel, which is exactly why the CDR design does not depend on the target covariate distribution or the data configuration. The simulation results are consistent with the theory, and the example is illustrative rather than confirmatory. The load-bearing weakness is not in the algebra but in the framing: the transportability of the conditional outcome mean is essential and is never stated as a formal assumption, while the abstract's 'regardless' language invites an overbroad reading. I do not think this warrants changing the verdict, because the body makes clear that Delta* is defined through the trial-conditional means and the exchangeability assumption is acknowledged by reference. A formal statement of the assumption and a small wording qualifier in the abstract would remove the ambiguity. The reader's weakest_assumption identified the same general area, but I partially disagree with the framing that trial variances are being used 'as if they applied in the target': for these data configurations the target outcome variance never enters the efficiency bound, since no target outcomes are observed. The condition that actually matters is equality of conditional means.","tokens_in":10564,"tokens_out":17106,"duration_ms":171138,"concrete_test":"Restate the needed condition formally as E[Y(a)|W=w] = m(a,w) for both populations, and independently re-derive the variance expression for the transportation influence function in Lemma 1 without assuming equal conditional outcome variances across populations. If the efficient influence function and its variance-minimizing p = sqrt(v1)/(sqrt(v1)+sqrt(v0)) are unchanged under this weaker condition, the invariance claim is confirmed to require only mean transportability; the remaining limitation is then purely the untestable mean-exchangeability assumption, which should be stated explicitly in the abstract or introduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theorem is sound conditional on the conditional outcome mean m(a,w) being the same in the trial and target populations; only then is Delta* = integral delta(w) dF*(w) the target-population ATE. In Section 2.1 this is invoked only as 'a certain exchangeability assumption' with citations, never stated as an explicit assumption. Because the abstract advertises maximal efficiency 'regardless of the target covariate distribution and the associated data configuration', a reader could reasonably infer that the design is robust to all population differences. In fact, if E[Y(a)|W] differs between populations, Delta* is not the target ATE and popt(W) is optimal for the wrong estimand. This is a scope limitation rather than an internal inconsistency: the calibrated-treatment-effect interpretation offered in Section 1 remains legitimate. One refinement to the reader's framing: popt uses trial-conditional variances v_a(W), not target variances, and this is correct because all outcome information in the three data configurations comes from the trial; the transportability that is actually load-bearing is mean exchangeability, not variance exchangeability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives optimal treatment allocation rules for estimating target-population effect measures from a randomized trial, in three data configurations: transportation, generalization, and post-stratification. The main result is that for estimands of the form ∆* = ∫δ(w)dF*(w) with δ(w)=m(1,w)−m(0,w), the optimal covariate-dependent randomization (CDR) design is p_opt(W)=√v_1(W)/(√v_1(W)+√v_0(W)), independently of the target covariate distribution and of which of the three data configurations supplies information about F*. The optimal covariate-independent randomization (CIR) design depends on F* through E{r(W)^2 v_a(W)} but has the same expression in all three configurations. The paper also extends the results to nonlinear contrasts, and reports a simulation study and an HIV trial example that match the theoretical predictions.","tokens_in":10788,"tokens_out":14600,"duration_ms":134743,"significance":"If the stated conditions are met, the invariance result is practically valuable: a single covariate-dependent randomization achieves the nonparametric efficiency bound for a range of target-population ATE estimands, simplifying trial design and analysis. The paper unifies three previously separate data configurations and the simulation confirms the predicted relative efficiencies. The main reservation is that the scope of the result is conditional on a mean-exchangeability assumption that the paper never states formally; the abstract's 'regardless' phrasing is accordingly stronger than what is proved. The variance-minimization derivations themselves are standard and consistent with the displayed influence functions.","major_comments":[{"comment":"The assumption that makes ∆* the target-population ATE is never stated formally. The text says only 'Under a certain exchangeability assumption (e.g., Zhang et al., 2016; Dahabreh et al., 2019; Colnet et al., 2024), ∆* is the ATE in the target population.' The abstract's claim of a design that is optimal 'regardless of the target covariate distribution and the associated data configuration' is therefore stronger than what is proved: without mean exchangeability, meaning m(a,w) identical in the trial and target populations, ∆* is a calibrated or adjusted contrast rather than the target causal effect, and p_opt is optimal for that contrast, not necessarily for the target ATE. Please state the mean-exchangeability assumption explicitly (variance exchangeability is not needed, because all outcome information comes from the trial), and qualify the abstract and theorem statements accordingly.","section":"Section 2.1"},{"comment":"The main results are stated without proof; Section 4 says technical proofs are in Supplementary Materials, which is not included in the arXiv version. Because the invariance of the optimal CDR design is the central contribution, the review version should include the supplement or at least the key variance decompositions that yield p_opt and π*_opt in each of the three configurations.","section":"Theorems 1-3"}],"minor_comments":[{"comment":"In Theorem 2 the second minimization clause says 'var{ψgen_cir(O)} is minimized uniquely by setting p equal to popt', and in Theorem 3 the second clause says 'var{ψps_cir(O)} is minimized uniquely by setting p = popt'; both should refer to the CDR influence functions ψgen_cdr and ψps_cdr.","section":"Theorems 2 and 3"},{"comment":"The text cites 'Capiello et al., 2021' but the reference list spells the name 'Cappiello'; please make the spelling consistent.","section":"Section 1 and references"},{"comment":"The reference list includes Ingall et al. (2004) and NINDS rt-PA Stroke Study Group (1995), but neither is cited in the body of the paper; either cite them or remove them.","section":"Section 4 and references"},{"comment":"There is a typo 'raltegraviror' in the first paragraph; it should be 'raltegravir'.","section":"Section 3.2"},{"comment":"The notation for the truncated normal is introduced as N(µ,σ^2;l,u) ∼ (X|l≤X≤u), but the simulation line writes N(0,0.75^2;-2,2) with the truncation limits in the same position; the definition should be stated in the order actually used to avoid ambiguity.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically sound and the central derivation is correct conditional on the usual exchangeability assumption. The main work for revision is formalizing that assumption and making the proofs or supplement available; both are local rather than conceptual problems. I do not see a need for a full re-review after these changes if the revision is handled carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper extends optimal Neyman-type allocation to target-population estimands and proves a clean invariance result for the ATE: the optimal covariate-dependent randomization design is the same regardless of target covariate distribution or whether the target information comes from transportation, generalization, or post-stratification. That result is new, correct, and worth having. The math is standard semiparametric variance minimization but executed cleanly, and the simulation matches the theory. I think the reader's ACCEPT verdict is right.\n\nWhat is genuinely new is the unified characterization through the density ratio r for CIR designs, and the stronger result that for ATE the optimal CDR design does not depend on F* at all. That is a real insight, not a minor variation on Zhang et al. (2023). The paper also notes that optimal CIR depends on F* but not on the data configuration, which usefully unifies three literatures. The extension to nonlinear contrasts in Section 2.5 is a nice compact addition.\n\nSoft spots, in order of importance. First, the paper never states the mean-exchangeability assumption formally as an assumption. Section 2.1 says \"Under a certain exchangeability assumption\" with citations and then moves on; Assumption 1 is only about support overlap. Since the abstract advertises a design that is optimal \"regardless of target covariate distribution and data configuration,\" a reader can easily infer robustness to all population differences. In fact, if the conditional outcome mean differs between trial and target, Delta* is not the target ATE and the optimal design is optimal for the wrong estimand. This is a scope limitation, not an internal flaw, but it should be stated as an explicit assumption. The paper does hedge in Section 1 by calling Delta* a calibrated effect, so a careful read is not badly misled; the abstract's rhetoric just overshoots. The stress-test refinement is right: popt uses trial conditional variances, and this is correct because all outcome information comes from the trial. The load-bearing assumption is mean exchangeability, not variance exchangeability.\n\nSecond, the proofs are deferred to Supplementary Materials not included in the arXiv version. For a theory paper that is a real accessibility problem for referees. Third, the simulation table lacks Monte Carlo standard errors, and Theorem 2 has a typo: the second minimization should refer to the CDR variance, not the CIR variance. These are minor relative to the central contribution.\n\nWho this is for: statisticians designing trials with external target populations, especially those working on transportability and generalizability. It deserves a serious referee. My recommendation: engage it, send it to review, and ask for the exchangeability assumption to be made explicit and the proofs to be posted.","headline":"The ATE invariance result is correct and new, but the paper should state the mean-exchangeability assumption explicitly rather than hiding it behind citations.","tokens_in":11297,"tokens_out":1958,"would_cite":true,"duration_ms":17082,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","62D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For estimating a target population's average treatment effect, one covariate-dependent randomization rule is optimal regardless of the target population or data setup.","keywords":["optimal treatment allocation","covariate-dependent randomization","average treatment effect","transportability","generalizability","post-stratification","efficiency bound","propensity score"],"falsifier":"Simulate a trial population and a target population with overlapping covariate supports but different conditional outcome variances, then estimate the variance of the efficient estimator under $p_{\\rm opt}(W)$ and under an allocation based on the target's own variances; if the latter achieves smaller variance for the target ATE, the claimed universality holds only under the exchangeability assumption, not as a universal property.","tokens_in":10367,"feed_emoji":"🎯","tokens_out":4419,"duration_ms":34855,"temperature":0.7,"pith_summary":"A randomized trial's treatment allocation can be tuned to estimate the effect of interest with maximum precision. This paper asks how to tune it when the estimand is the average treatment effect (ATE) in a target population that overlaps with, but differs from, the trial population. It proves that a single covariate-dependent randomization rule — assign each patient to treatment with probability proportional to the conditional standard deviation of the outcome under that treatment — maximizes asymptotic efficiency for the target ATE, regardless of the target covariate distribution and regardless of whether the target information comes from transportation, generalization, or post-stratification. The same rule is already optimal for the trial population's own ATE. If correct, trial designers can pre-specify one allocation that serves the trial estimand and any externally defined target estimand simultaneously.","feed_headline":"One allocation rule is optimal for every target population","feed_subtitle":"The same covariate-dependent randomization maximizes efficiency for transported, generalized, and post-stratified ATEs.","key_machinery":"The engine is the semiparametric efficiency bound: for each data configuration (independent target cohort, nested target cohort, or known stratum weights) the paper writes the efficient influence function for the target ATE, then minimizes its variance over the design's propensity score. For CDR the influence functions under all three configurations share the same dependence on the propensity score, and differentiating the variance gives a pointwise first-order condition whose unique solution is the variance-proportional allocation $p_{\\rm opt}(W)$. For CIR the same variance calculation yields a scalar equation solved by the density-ratio-weighted allocation $\\pi^*_{\\rm opt}$. The density ratio $r(W)=dF^*/dF(W)$ is the object that carries the target population's influence for CIR, and it cancels entirely from the CDR solution.","core_discovery":"Under an exchangeability assumption that the conditional outcome mechanism is shared between trial and target populations, the paper derives the efficient influence function for the target ATE in each of three data configurations and minimizes its variance over designs. The minimizer for covariate-dependent randomization (CDR) is uniquely $p_{\\rm opt}(W)=\\sqrt{v_1(W)}/(\\sqrt{v_1(W)}+\\sqrt{v_0(W)})$, where $v_a(W)$ is the conditional variance of the potential outcome under treatment $a$. This design does not depend on the target covariate distribution $F^*$ or on which of the three configurations supplies information about $F^*$; it is the Neyman allocation applied within each covariate stratum. The optimal covariate-independent randomization (CIR) design, by contrast, is $\\pi^*_{\\rm opt}=[E\\{r(W)^2 v_1(W)\\}]^{1/2}/([E\\{r(W)^2 v_1(W)\\}]^{1/2}+[E\\{r(W)^2 v_0(W)\\}]^{1/2})$, which depends on the target through the density ratio $r=dF^*/dF$ but not on the data configuration. For nonlinear effect measures of the form $g(\\mu^*_1)-g(\\mu^*_0)$, the optimal CDR allocation depends on the target only through the two target means, not through the density ratio.","pith_inferences":["The result suggests a practical workflow: estimate conditional outcome variances from pilot or historical data, pre-specify $p_{\\rm opt}(W)$, and the same trial design supports a family of hypothetical target populations.","If outcome variances are misspecified, the efficiency loss under CDR is likely second-order, so the design may be robust to moderate misspecification, though the paper does not quantify this.","The invariance may extend to multi-arm trials or to estimands defined by weighted averages of subgroup effects; testing that extension would require deriving the corresponding influence functions.","A testable implication is that registry-based post-stratification with published summary weights can use the trial-optimal CDR design without re-optimizing, which the HIV example illustrates."],"forward_implications":["A trial can be designed with covariate-dependent randomization using only trial-population outcome variance estimates, and the same design is asymptotically optimal for the trial ATE and for any transportable target ATE.","Optimal covariate-independent randomization does depend on the target distribution through the density ratio, so a single fixed randomization fraction cannot serve all target populations.","For ATE estimation, CDR strictly dominates CIR in efficiency and removes the need to know the target covariate distribution when choosing the design.","For nonlinear contrasts such as risk differences on the log or logit scale, the optimal CDR depends on the target only through the two marginal potential-outcome means, so the density ratio still does not enter.","The invariance result holds identically under transportation, generalization, and post-stratification, so the data configuration does not affect the design choice."],"supporting_citations":[{"why":"Derives the optimal CDR design $p_{\\rm opt}(W)$ for the trial-population ATE, which the paper extends and shows invariant to the target population.","marker":"Zhang et al. (2023)"},{"why":"Supplies the efficient influence function for transportation under CIR, which the paper adapts to CDR and minimizes.","marker":"Zhang et al. (2016)"},{"why":"Provides the efficient influence function for generalization under CIR, the second data configuration the paper unifies.","marker":"Dahabreh et al. (2019)"},{"why":"Origin of the variance-proportional allocation that the optimal CDR design applies within covariate strata.","marker":"Neyman (1934)"},{"why":"Shows machine-learning based estimators can attain efficiency for transported effects, justifying the efficiency-bound objective.","marker":"Capiello et al., 2021"},{"why":"Cited for the exchangeability assumption under which the target estimand equals a causal ATE.","marker":"Colnet et al., 2024"}],"fun_headline_variants":["Optimal allocation independent of target population","One randomization rule for all target populations","Population-agnostic optimal treatment design","Efficient allocation robust to population shifts","Same optimal design for any target covariate mix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trial and target populations must share the same conditional outcome distribution — same mean and same variance given covariates — so that the trial's variance functions apply to the target estimand.","fun_headline_variants_meta":{"raw":{"variants":["Optimal allocation independent of target population","One randomization rule for all target populations","Population-agnostic optimal treatment design","Efficient allocation robust to population shifts","Same optimal design for any target covariate mix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2765,"prompt_tokens":998,"completion_tokens":1767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":1714}},"tokens_in":614,"tokens_out":1767,"duration_ms":11499,"temperature":1.0,"reasoning_tokens":1714,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:09:54.646701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a trial population and a target population with overlapping covariate supports but different conditional outcome variances, then estimate the variance of the efficient estimator under $p_{\\rm opt}(W)$ and under an allocation based on the target's own variances; if the latter achieves smaller variance for the target ATE, the claimed universality holds only under the exchangeability assumption, not as a universal property.","supporting_citations":[{"cited_title":"and Hu, Z","cited_arxiv_id":null,"evidence_quote":"Supplies the efficient influence function for transportation under CIR, which the paper adapts to CDR and minimizes."},{"cited_title":"and Hernan, M.A","cited_arxiv_id":null,"evidence_quote":"Provides the efficient influence function for generalization under CIR, the second data configuration the paper unifies."},{"cited_title":"(2024) Causal inference methods for combining randomized trials and observational studies: A review","cited_arxiv_id":null,"evidence_quote":"Cited for the exchangeability assumption under which the target estimand equals a causal ATE."}],"review_version":1}