{"id":"f35a6444-1f6c-460d-8b41-4a65dfa1c464","arxiv_id":"2412.12380","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper argues that the ICH E9(R1) estimands framework and causal inference define the same kinds of target effects in clinical trials, so they should be used as complementary tools.","lead":"This paper compares two ways of defining what a clinical trial measures: the ICH E9(R1) estimands framework used in drug development and the potential outcomes framework from causal inference. It argues they are complementary views of the same goal, each with different strengths in clarity and mathematical precision.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Complementarity is demonstrated for four of the five ICH strategies, but the unillustrated 'while on treatment' strategy may not map to a causal estimand, so Section 6.1's identification of intercurrent events with mediators or time-varying treatments is broader than the examples support.","rationale":"Good-faith reading: the paper is a conceptual synthesis, not a mathematical theorem, and its examples are carefully chosen and mostly convincing. The treatment-policy, hypothetical, composite, and principal-stratum translations in Section 5 are correct in substance; the authors also deserve credit for flagging ambiguities in ICH, including the treatment-policy label, the principal-stratum misinterpretation, and the arbitrary penalty value in composite strategies. The load-bearing point is the mapping in Section 6.1. If IEs are only sometimes 'mediators or time-varying treatments,' then the frameworks may diverge for the remaining cases, and the conclusion's unqualified 'both can be used' overstates the demonstrated scope. The missing 'while on treatment' strategy is the clearest test case because the paper defines it in Section 3 but never revisits it in Section 5. The concrete test settles whether an actual divergence exists. If the translation fails, the fix is a qualification rather than a rejection, hence CONDITIONAL rather than REJECT.","tokens_in":13804,"tokens_out":9170,"duration_ms":88408,"concrete_test":"Take the 'while on treatment' strategy from Section 3 with death as the intercurrent event in a time-to-event trial, and attempt to write the ICH estimand as a comparison of potential outcomes Y(a) in the style of Section 5 without conditioning on post-treatment survival or adding principal-stratum assumptions. If no such causal estimand exists without these extra assumptions, the complementarity claim in Section 7 should be explicitly restricted to the four illustrated strategies; if one does exist, state it and verify that it reproduces the ICH definition.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that the ICH and causal-inference frameworks can both quantify the same population-level target effect—depends on the mapping proposed in Section 6.1: an intercurrent event (IE) is treated as 'post-baseline mediator or time varying treatment.' The paper illustrates this mapping for treatment policy, hypothetical, composite, and principal stratum strategies, but not for the fifth ICH strategy, 'while on treatment.' Under that strategy the outcome is truncated at the occurrence of a post-baseline event; when the event is death or treatment discontinuation, no standard potential-outcome quantity corresponds to the truncated mean without additional structural assumptions such as principal stratification or a fully specified dynamic regime. Section 3 itself calls the strategy's label 'arguably ambiguous,' yet the conclusion asserts broadly that both approaches 'can be used to quantify a population-based summary of a target effect.' The examples therefore support complementarity only for a subset of ICH strategies, and the broader claim is asserted rather than established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares the ICH E9(R1) estimands framework with the causal inference framework, using three clinical trial examples (diabetes, nasal polyps, COVID-19) to show how each can define estimands that address intercurrent events. The authors argue that, despite differences in form and precision, the two paradigms are complementary: the ICH framework provides a structured, descriptive approach for communicating trial objectives, while causal inference offers mathematical precision and explicit assumptions via potential outcomes and DAGs. They conclude that both can quantify a population-based summary of a target effect and encourage practitioners to use both frameworks in combination.","tokens_in":14023,"tokens_out":7725,"duration_ms":65182,"significance":"If the central claim holds, the paper provides a valuable didactic bridge between two important statistical frameworks, helping trialists and statisticians understand when and how ICH E9(R1) estimands correspond to causal estimands. Its strengths are the worked examples covering treatment policy, hypothetical, composite, and principal stratum strategies; the explicit acknowledgment of ambiguity in the ICH guidance; and the balanced discussion of limitations such as the untestable assumptions for principal strata and the arbitrariness of composite penalty values. The paper does not contain new mathematical results, but it offers a clear and practical framework for cross-communication, which is of genuine use to the clinical trials community.","major_comments":[{"comment":"The paper defines five intercurrent-event strategies in §3 but never revisits the \"while on treatment\" strategy in any of the examples, which cover only treatment policy, hypothetical, composite, and principal stratum strategies. The conclusion in §7 asserts broadly that both paradigms \"can be used to quantify a population-based summary of a target effect on an outcome for a given population,\" and §6.1 states that an IE is conceptually the same as a post-baseline mediator or time-varying treatment. Without a discussion of how the while-on-treatment strategy maps to a causal estimand—for example, via dynamic treatment regimes or principal stratification—the complementarity claim is wider than the demonstrated evidence. Please add such a discussion or explicitly qualify the conclusion to the strategies actually illustrated.","section":"§3, §6.1, §7"},{"comment":"The proposed mapping between intercurrent events and mediators or time-varying treatments is presented as \"reasonably intuitive\" and general, but it does not obviously accommodate death as an intercurrent event, which the paper itself lists in §3 as an important IE in some trials. When death occurs before the outcome measurement, the outcome is undefined in the usual sense, and neither a mediator nor a time-varying treatment description applies without additional structural assumptions or a composite or principal-stratum formulation. The paper should acknowledge this limitation and either provide a worked example involving death or temper the general claim that all IEs can be identified with causal-inference constructs.","section":"§3, §6.1"}],"minor_comments":[{"comment":"The individual causal effect is written as Y(a=1) – Y(a=1) and later as E[Y(a=1) – Y(a=1)]; these should be Y(a=1) – Y(a=0) and E[Y(a=1) – Y(a=0)]. This appears to be a simple typo, but it occurs in the central definition of the causal effect.","section":"§4.1"},{"comment":"The notation E[Y(a=1, D, R)] is nonstandard because D and R are random variables, not fixed values. Consider writing E[Y(a=1)] directly or explicitly defining Y(a, d, r) with fixed d and r and using language such as \"with D and R taking their natural values under each assignment.\"","section":"§5.1.2"},{"comment":"The claim that the ICH estimand is \"likely to be more accessible to clinical trialists without formal causal and mathematical training\" is an empirical assertion that is not supported by evidence in the manuscript. Please either cite supporting literature or present this as the authors' opinion rather than an established fact.","section":"§6.1"},{"comment":"The phrase \"assumptions about the casual relationships\" should read \"assumptions about the causal relationships.\"","section":"§5.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful and well-written didactic contribution that fits the journal's scope. The main issue is the overgeneralization of the complementarity claim without treating the 'while on treatment' strategy or death as an intercurrent event, both of which are listed in the paper itself. A focused revision addressing these points would make the argument fully convincing for the broad audience the paper targets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work in drug-development statistics or teach estimands. It's a conceptual paper, no new methods, but it does something useful: it takes three real trials and writes the same target effect in both ICH E9(R1) and potential-outcomes notation side by side. The worked comparisons are the contribution, and they are mostly careful and correct. The COVID-19 example does a clean job separating the hypothetical effect from the principal-stratum effect, a distinction that still confuses people. The commentary on ambiguous ICH language—the 'treatment policy' misnomer for assignment effects, and the principal-stratum wording that invites a post-hoc-subgroup misreading—is practical and fair. The only math error I found is a typo in Section 4.1, where the expected difference is printed as E[Y(a=1) - Y(a=1)] instead of E[Y(a=1)] - E[Y(a=0)].\n\nThe main soft spot is the mapping claim in Section 6.1. The paper says it is 'reasonably intuitive' to treat an intercurrent event as a post-baseline mediator or time-varying treatment, but the examples only demonstrate this for four of the five ICH strategies. The fifth, 'while on treatment,' truncates the outcome at a post-baseline event; when that event is death, no standard potential-outcome quantity corresponds to the truncated mean without extra structural assumptions. The paper itself calls the label 'arguably ambiguous' in Section 3, so the conclusion that both frameworks quantify the same population-based summaries is asserted more broadly than the examples establish. That's a gap a paragraph on the fifth strategy would fix, not a reason to reject. The accessibility claim—ICH is easier for non-mathematicians—is also asserted without evidence; plausible, but untested.\n\nOn novelty: the paper cites Lipkovich et al. 2020, Han and Zhou 2023, and Scharfstein's slides, so the linkage idea is not new. The new content is the worked examples and the specific critique of ICH ambiguity. The self-citations are background estimation references, not premises that force the conclusion.\n\nBottom line: modest, honest, clearly argued paper for trialist statisticians, protocol writers, and teachers of estimands. It deserves a serious referee. I would ask for the 'while on treatment' caveat and a softer hedge on accessibility, then accept.","headline":"Clear conceptual paper showing ICH E9(R1) and causal estimands map onto each other for four of the five intercurrent-event strategies; the unillustrated 'while on treatment' case is a revision-level gap, not a rejection.","tokens_in":745,"tokens_out":2370,"would_cite":true,"duration_ms":57579,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The ICH E9(R1) estimand framework and causal inference define the same target effect and should be used together.","keywords":["ICH E9(R1)","estimands","causal inference","potential outcomes","intercurrent events","clinical trials","treatment policy","principal stratum"],"falsifier":"Take a real trial with an intercurrent event that is not a well-defined intervention, such as death as a competing risk for a quality-of-life endpoint, and try to write the ICH treatment-policy and hypothetical estimands as causal estimands in the paper's notation; if the formulations diverge or one is inexpressible, the claimed complementarity fails for that class of events.","tokens_in":13631,"feed_emoji":"🎯","tokens_out":6794,"duration_ms":54397,"temperature":0.7,"pith_summary":"This paper argues that the ICH E9(R1) estimands framework and the causal inference framework are not rivals but two ways of stating the same thing: a population-based summary of a treatment effect on an outcome. Using three real trial settings (diabetes, nasal polyps, COVID-19 vaccine), the authors show that each ICH estimand can be rewritten as a potential-outcomes expression, and that the ICH intercurrent-event strategies correspond to causal concepts such as effect of assignment, intervening on post-baseline variables, composite outcomes, and principal strata. If the authors are right, trialists can keep the ICH framework's descriptive table for communication while borrowing causal inference's mathematical precision and explicit assumptions for specification. That combination would make it easier to say precisely what a trial is estimating and why.","feed_headline":"Two estimand schools target one effect: complementary, not competing","feed_subtitle":"Pair ICH estimand structure with causal precision to say exactly what a trial is estimating.","key_machinery":"The mapping between ICH intercurrent-event (IE) strategies and causal inference concepts is the load-bearing object. The paper identifies treatment policy with the effect of assignment, hypothetical with an intervention that sets post-baseline variables to specified values, composite with a redefined outcome that incorporates the event (with a penalty value), while-on-treatment with measuring the outcome only until the event, and principal stratum with the effect in the subpopulation that would or would not experience the event under both treatments. This correspondence is what allows the paper to claim that a single target effect can be expressed in either framework, and that the frameworks are complementary.","core_discovery":"The central claim is that 'estimand' in ICH E9(R1) and 'estimand' in causal inference denote essentially the same target: a comparison, in a specified population, of potential outcomes under specified treatment conditions. The paper demonstrates this by formulating the same clinical questions in both languages for three trials: the treatment-policy strategy matches the causal estimand that leaves post-baseline variables unintervened on; the hypothetical strategy matches a causal estimand that intervenes to set those variables (e.g., no discontinuation, no rescue); the composite strategy matches redefining the outcome as a mixture of the natural outcome and a penalty value; and the principal-stratum strategy matches conditioning on potential event status. The paper concludes that the two frameworks are complementary and encourages practitioners familiar with one to learn the other.","pith_inferences":["The mapping works most cleanly for intercurrent events that can be framed as interventions; for terminal events like death, the causal language would need new machinery, suggesting the complementarity may be narrower than stated.","The paper's accessibility claim is untested; a survey asking clinical trialists to interpret both formulations would provide direct evidence for or against it.","A practical by-product could be a translation guide that pairs each ICH strategy with its causal estimand notation, going beyond the three examples given here."],"forward_implications":["A trial team can present the ICH five-component table to stakeholders and the corresponding potential-outcome expression to statisticians, keeping both audiences clear.","Adding a causal graph to an ICH estimand makes the assumed relationships between treatment, intercurrent events, and outcome explicit rather than implicit.","The hypothetical strategy, read causally, is an effect under a specified intervention on adherence and rescue use, directly linking it to per-protocol effect definitions.","Principal-stratum estimands, the paper notes, are only estimable under strong and untestable assumptions, a point the ICH text does not spell out.","Adopting both vocabularies should reduce the confusion that arises when the single word 'estimand' is used in two partially overlapping senses."],"supporting_citations":[{"why":"Supplies the ICH E9(R1) estimand framework and its five components, the object being compared.","marker":"[1]"},{"why":"Defines the potential-outcomes language and causal estimands used for the comparison.","marker":"[6]"},{"why":"Provides the origin of the estimand concept in missing-data guidance.","marker":"[10]"},{"why":"Shows how hypothetical-strategy estimands unify causal inference and missing-data methods.","marker":"[25]"},{"why":"Discusses the arbitrariness of penalty values in composite estimands.","marker":"[33]"},{"why":"Raises the critiques of principal-stratum estimands that the paper weighs.","marker":"[34]"},{"why":"Represents prior attempts to link intercurrent-event strategies to causal effects.","marker":"[36]"},{"why":"Defines the per-protocol causal effect that a hypothetical strategy corresponds to.","marker":"[40]"}],"fun_headline_variants":["Two frameworks, one estimand: complementary, not competing","Estimand frameworks: ICH and causal inference are allies","ICH and causal estimands: same target, different tools","One estimand, two languages: ICH meets causal inference","Causal clarity meets ICH structure for precise trial estimands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The complementarity claim rests on the assumption that every intercurrent event can be mapped onto a causal concept—an intervenable post-baseline variable or a principal stratum—and for some events, such as death, that mapping may not be well defined.","fun_headline_variants_meta":{"raw":{"variants":["Two frameworks, one estimand: complementary, not competing","Estimand frameworks: ICH and causal inference are allies","ICH and causal estimands: same target, different tools","One estimand, two languages: ICH meets causal inference","Causal clarity meets ICH structure for precise trial estimands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1483,"prompt_tokens":919,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":481}},"tokens_in":535,"tokens_out":564,"duration_ms":5252,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:07:18.163579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real trial with an intercurrent event that is not a well-defined intervention, such as death as a competing risk for a quality-of-life endpoint, and try to write the ICH treatment-policy and hypothetical estimands as causal estimands in the paper's notation; if the formulations diverge or one is inexpressible, the claimed complementarity fails for that class of events.","supporting_citations":[{"cited_title":"E9(R1) (2019) Addendum on Estimands and Sensitivity Analysis in Clinical Trials","cited_arxiv_id":null,"evidence_quote":"Supplies the ICH E9(R1) estimand framework and its five components, the object being compared."},{"cited_title":"Causal inference: What if","cited_arxiv_id":null,"evidence_quote":"Defines the potential-outcomes language and causal estimands used for the comparison."},{"cited_title":"https://www.ncbi.nlm.nih.gov/books/NBK209902/; 2010","cited_arxiv_id":null,"evidence_quote":"Provides the origin of the estimand concept in missing-data guidance."},{"cited_title":"Strategies for composite estimands in confirmatory clinical trials: Examples from trials in nasal polyps and steroid reduction","cited_arxiv_id":null,"evidence_quote":"Discusses the arbitrariness of penalty values in composite estimands."},{"cited_title":"Translating questions to estimands in randomized clinical trials with intercurrent events","cited_arxiv_id":null,"evidence_quote":"Raises the critiques of principal-stratum estimands that the paper weighs."},{"cited_title":"Per-Protocol Analyses of Pragmatic Trials","cited_arxiv_id":null,"evidence_quote":"Defines the per-protocol causal effect that a hypothetical strategy corresponds to."}],"review_version":1}