{"id":"e5886633-e9a4-4ca9-908d-234a3f526648","arxiv_id":"2608.11108","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"For studies of pre-pregnancy medication changes in early pregnancy, time zero should be the first prenatal healthcare visit, with clone-censor-weighting or sequential trials to avoid bias.","lead":"This paper proposes how to design real-world data studies of medication changes in early pregnancy, anchoring time zero at the first prenatal visit. It recommends clone-censor-weighting and sequential trials to reduce immortal time and selection bias.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Time zero anchored at first pregnancy-related encounter assumes no treatment decision occurred before that encounter; if early changers exist, the design conditions on survival and misclassifies pre-encounter switches, weakening the central claim.","rationale":"The reader's weakest assumption concerns measurement error: prescription fills within a grace period may not reflect true use because of stockpiling, leftover medication, or non-adherence. My concern is distinct but related: even with perfectly accurate fill records, the chosen time zero may be later than the actual treatment decision. The paper's strongest claim is that time zero should be anchored on a healthcare encounter, but this only works if the first relevant encounter is also the first opportunity for a decision. The authors explicitly target pregnant people with T2DM 'taking metformin at the time of entry into prenatal care,' and their Figure 3 illustrates a woman who has her first prenatal visit at week 11 and then is cloned into strategies. If many women in real data change their medication before that visit, the strategy definitions based on post-visit fills will misclassify them, and the analytic approaches that 'minimize immortal time bias' will not address the immortal time accumulated before the encounter. The paper acknowledges that conception is not a feasible time zero, but it does not discuss treatment changes that occur before the first pregnancy-coded visit, such as changes made after a home pregnancy test or at a non-obstetric visit. This is a real gap because the clinical question is about early pregnancy decisions, and those decisions are not all made at prenatal visits. A conditional acceptance would ask the authors to clarify the target population (only those with no pre-encounter change) or to show empirically that early changes are rare. This is a substantive but not fatal issue; the core framework is useful when the decision point is truly the encounter.","tokens_in":11019,"tokens_out":7519,"duration_ms":73000,"concrete_test":"In a linked claims/EMR cohort with identified pregnancy episodes, identify women with pregestational T2DM on metformin and chart, for each pregnancy, the timing of the first pregnancy-related encounter and the timing of any metformin discontinuation or insulin initiation (fill dates, days' supply, and encounter codes). Compare the distribution of treatment-change dates relative to the first prenatal visit.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central recommendation (Section 3: 'time zero should be anchored on a healthcare encounter'; 'The first such visit can be used as the first possible time at which a person might be enrolled') presupposes that the first pregnancy-related encounter is also the first time at which a treatment change could occur. In routinely collected data, this is often false. A woman with pregestational T2DM may discontinue metformin the day after a positive home pregnancy test, or her primary-care physician may switch her to insulin before a prenatal referral is coded. Such decisions occur before the first prenatal visit and are invisible to the proposed time zero. Cloning and censoring from the first prenatal visit then (i) restricts the study population to pregnancies that survived to prenatal care without a prior treatment change, inducing selection bias for the clinically relevant question, and (ii) misclassifies a woman who already switched before baseline as following a 'continue' or 'post-visit switch' strategy, because the observed post-visit fills are interpreted as strategy initiation after time zero. The paper's Figure 2 and Section 4 discuss the need to avoid looking into the future for decisions after time zero, but they do not address left-censoring of decisions before time zero. The grace-period and clone-censor-weight designs only fix immortal time that starts at the chosen encounter; they cannot fix immortal time that accrued before it. Unless the target population is explicitly restricted to those who have made no treatment change before the first encounter, or data include non-prenatal encounters and pregnancy testing dates, the recommended anchoring does not deliver the promised alignment with real-world decision timing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a methods/guidance paper on designing perinatal pharmacoepidemiology studies that evaluate changes to pregestational medication regimens, using type 2 diabetes mellitus treated with metformin as the running example. The authors argue that for such questions, time zero should be anchored at healthcare encounters (e.g., the first prenatal visit) rather than at conception; that treatment strategies should be defined in relation to existing treatment (continuation, adding, switching, discontinuation); and that clone-censor-weighting or sequential trial designs should be used to avoid immortal person time and conditioning on future events. They also review algorithms for identifying and dating pregnancies in administrative data, discuss how prescription fills can be used to operationalize strategies within a grace period, and enumerate limitations of claims-based measurement of treatment strategies.","tokens_in":11259,"tokens_out":10016,"duration_ms":90466,"significance":"If the proposed design principles are adopted, they would move the field from exposure-oriented 'first-trimester exposed versus unexposed' comparisons toward decision-oriented comparative effectiveness questions that match the timing of clinical care, which is a substantive contribution. The paper is careful to acknowledge residual within-trial immortal time in sequential trials, confounding that is not addressed by design, and measurement limitations of prescription fills. The protocol table and the illustrative figures provide concrete starting points for implementation, and the paper explicitly identifies falsifiable design assumptions (e.g., the grace period and the use of fills to define strategies). The manuscript does not present empirical analyses, simulations, or code, but as a design framework that is not a deficiency.","major_comments":[{"comment":"The central recommendation that 'time zero should be anchored on a healthcare encounter' and that 'The first such visit can be used as the first possible time at which a person might be enrolled' does not account for treatment decisions made before the first observed pregnancy-related encounter. A pregnant person with T2DM may stop metformin or start insulin immediately after a positive home pregnancy test, or at a primary care visit that is not coded as prenatal care. Such decisions occur before time zero and are not addressed by the cloning/censoring machinery: if eligibility is operationalized as 'using metformin at the first prenatal visit,' early switchers are excluded and the estimand becomes conditional on survival without treatment change up to that visit; if eligibility is instead inferred only from pre-pregnancy fills, early switchers can be misclassified as 'continuers' because the baseline strategy is read from fills observed after the decision has already been made. The manuscript should either explicitly restrict its scope to decisions made at prenatal or other healthcare encounters, or discuss how earlier decision points could be identified (e.g., pregnancy-test-related visits or primary care visits) and what sensitivity analyses would be needed when the timing of the decision is unknown.","section":"Section 3, 'Aligning time zero with relevant time points for treatment decisions'; Figure 1"}],"minor_comments":[{"comment":"The Causal Contrast row is confusingly formatted: the target trial column reads 'Intention-to-treat effect; per-protocol effect' and the emulation column also lists 'Per-protocol effect.' Please clarify whether both ITT and per-protocol effects are estimands, or only the per-protocol effect.","section":"Table, Causal Contrast row"},{"comment":"The sentence 'residual immortal person time bias may still be present within each trial, since initiators can fill a prescription at any time within the trial interval while non-initiators must complete the full interval without filling a prescription' is cited to reference 35, which concerns gestational age at arrest of development; this reference does not support the statement about sequential trials and should be replaced with a methods citation on sequential trial emulation (e.g., Caniglia et al. 2023 or a standard text).","section":"Section 4, references"},{"comment":"The author affiliation numbering is jumbled, with duplicate '4' labels for the University of British Columbia and McGill University; please correct the affiliation list.","section":"Author affiliations"},{"comment":"The statement that for Pregnancy 2 'no data would be recorded' and researchers 'would not be aware of its existence until the abortion procedure occurred' is internally inconsistent, since the abortion procedure itself is a healthcare encounter; please rephrase to 'no data would be recorded before the abortion claim.'","section":"Section 1, Figure 1 discussion"},{"comment":"The phrase 'The first such visit can be used as the first possible time at which a person might be enrolled' should be qualified as 'the first pregnancy-related visit observable in the data at which eligibility can be assessed,' because earlier healthcare encounters (e.g., primary care visits with a pregnancy test) may exist and may already be decision points.","section":"Section 3, page 9"},{"comment":"The discussion of stockpiling, leftover medication, and provider samples is useful; consider adding a sentence recommending quantitative bias analysis to explore how such misclassification of strategy adherence would affect estimates, paralleling the sensitivity-analysis recommendation already made for selection bias in the Discussion.","section":"Section 2, measurement of strategies"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The paper is a credible and clearly written methods-guidance contribution that fits the journal's scope. The main substantive issue is the scope of the time-zero recommendation with respect to treatment decisions that occur before the first observed pregnancy-related encounter; this is fixable through a scoping clarification and a discussion of sensitivity analyses. The self-citations are contextual and not load-bearing, and I see no grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, clearly-written guidance paper that fills a real gap. Prior target trial emulation work in pregnancy focused on point treatments like vaccines or antibiotics; this one tackles changes to pregestational treatment regimes — continuation, add-on, switch, discontinuation — using T2DM/metformin as a running example. That is the genuinely new piece, and it is handled well. The clone-censor-weight and sequential trials sections are accurate, the figures are genuinely helpful, and the paper is honest about residual problems: within-trial immortal time, confounding that design cannot fix, and the limits of claims data. The target trial table is a useful reference for anyone designing such a study.\n\nThe main soft spot is exactly what the stress-test flagged, though it is not fatal. Anchoring time zero at the first prenatal encounter implicitly assumes no treatment change occurred before that encounter. A woman who switches to insulin the day after a positive home test, before any coded visit, is either excluded or misclassified as following a baseline strategy. The paper does define the target population as taking metformin at entry into prenatal care, which partly answers the concern, but that restriction is buried. The authors should state more prominently that the estimand is for the population that reaches the first encounter unchanged, and they should suggest sensitivity analyses or alternative time zeroes (any encounter, pregnancy testing dates) for questions about earlier changes. The grace period of 45 days is also a bit arbitrary; the paper acknowledges this but does not discuss how to choose it.\n\nThe citation pattern is fine. The self-citations (Latour, Chiodo, Wood) provide background context and are not load-bearing. The paper is a guidance piece, so absence of data is not a flaw. For perinatal pharmacoepidemiologists and methodologists, it deserves a serious referee and likely a place in the literature. I would accept after minor revisions; the pre-visit left-censoring discussion should be more explicit.","headline":"Useful methods guidance for a neglected design question; the pre-visit left-censoring issue is real but within scope if the target population is made explicit.","tokens_in":11856,"tokens_out":2001,"would_cite":true,"duration_ms":18698,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For studies of medication changes in early pregnancy, follow-up should begin at the healthcare encounter where the treatment decision is made—usually the first prenatal visit—not at conception.","keywords":["pregnancy","pharmacoepidemiology","target trial emulation","time zero","immortal time bias","clone-censor-weighting","sequential trials","administrative health data"],"falsifier":"A head-to-head emulation of the same treatment-change question in the same administrative data, run once with conception as time zero and once with the first prenatal encounter as time zero, would settle the claim if the two estimates diverge in the direction and magnitude predicted by immortal time bias; if they do not diverge, re-anchoring time zero changes nothing.","tokens_in":1558,"feed_emoji":"🤰","tokens_out":1834,"duration_ms":83609,"temperature":0.7,"pith_summary":"The paper argues that studies of medication changes during early pregnancy should not treat conception as the natural start of follow-up. Instead, follow-up should begin at the healthcare encounter where the treatment decision actually happens, typically the first prenatal visit, and the strategies being compared should be defined relative to the medication a person already takes: continue, add, switch, or discontinue. The reason is that anchoring at conception or defining strategies by looking at future prescription fills can create immortal person time and selection bias. Using type 2 diabetes as a worked example, the paper walks through identifying pregnancies in administrative data, choosing a time zero, and analyzing with clone-censor-weighting or sequential trials. If the approach is right, perinatal pharmacoepidemiology studies would answer the question a patient faces at the first prenatal visit rather than a question about exposures dated from conception.","feed_headline":"Anchor pregnancy drug studies at prenatal care, not conception","feed_subtitle":"How to avoid immortal time bias when comparing continuing, adding, or switching medications in early pregnancy.","key_machinery":"The central mechanism is the encounter-anchored time zero: starting follow-up at a prenatal visit or other healthcare encounter where a treatment decision is actually made, rather than at conception. The argument is carried by two analytic devices—clone-censor-weighting, in which each observation is cloned once per strategy and clones whose observed fills deviate from their assigned strategy are censored and reweighted, and sequential trials, in which non-initiators re-enter eligibility at later intervals. Both require defining strategies relative to the pre-pregnancy regimen and fixing a grace period, illustrated as 45 days, so that no future prescription fills are used to decide which strategy a person followed.","core_discovery":"The central claim is that for research questions about changing pregestational medication regimens, the time zero of a target trial emulation should be a healthcare encounter, not conception. Conception is intuitively appealing but does not align with real decision points, because pregnancy recognition and entry into care happen weeks later. The paper specifies treatment strategies relative to the regimen already in use—continuation, adding, switching, and discontinuation—and shows that with a prespecified grace period, encounter-anchored designs can be analyzed with clone-censor-weighting or sequential trials to avoid immortal person time and conditioning on future events. The worked example is a hypothetical trial comparing metformin continuation, metformin plus insulin, and switching to insulin among people with pregestational type 2 diabetes at their first prenatal visit.","pith_inferences":["Editorial inference: the same encounter-anchored logic extends beyond diabetes to any chronic pregestational medication—antidepressants, antihypertensives, immunosuppressants—and beyond pregnancy to any setting where treatment decisions occur at discrete clinical visits.","Editorial inference: running the same research question on the same database with conception-anchored and encounter-anchored time zero would quantify how much immortal time bias actually shifts estimates in perinatal pharmacoepidemiology.","Editorial inference: formalizing a data-generating model in which fills misrepresent true use through stockpiling, leftover medication, or non-adherence would yield bounds or bias-calibrated estimates for the clone-censor-weight estimator, addressing the paper's acknowledged uncertainty about fill accuracy.","Editorial inference: the grace-period length is itself a design choice that may interact with care access; when prenatal visits are infrequent, a fixed 45-day window could misclassify strategies, so adaptive grace periods defined by the next encounter are worth testing."],"forward_implications":["Studies that anchor at conception and simply compare exposed versus unexposed pregnancies should be replaced or supplemented by encounter-anchored designs when the question is whether to continue, add, switch, or stop a chronic medication.","Pregnancy identification algorithms based on delivery or live birth records condition on a future event; anchoring at the first pregnancy-related encounter restores the correct risk set of pregnancies at the decision point.","Using a prespecified grace period to classify strategies reduces the chance that researchers look into future prescription fills to determine a person's strategy.","Clone-censor-weighting and sequential trials both fit the pregnancy setting, with sequential trials best suited to initiation-versus-no-initiation comparisons and clone-censor-weighting suited to multi-strategy decisions.","Clinical trial templates from pregnancy trials, such as the metformin-plus-insulin example, can directly inform the design of emulated trials in administrative data."],"supporting_citations":[{"why":"Shows that the method used to identify pregnancies is itself a source of bias, motivating the paper's emphasis on who can be enrolled at an encounter.","marker":"[4]"},{"why":"A recent randomized pregnancy trial that supplies the clinical template for the hypothetical metformin-plus-insulin trial.","marker":"[11]"},{"why":"Systematic review of time-related biases in perinatal pharmacoepidemiology, the problem these designs are meant to solve.","marker":"[36]"},{"why":"Provides the target trial emulation and sequential-trial template used to avoid immortal time in pregnancy settings.","marker":"[37]"},{"why":"Describes clone-censor-weighting in detail, the paper's primary analytic adaptation.","marker":"[38]"},{"why":"Reflects on clone-censor-weight and trial emulation, another basis for the proposed methods.","marker":"[39]"},{"why":"Example of target trial emulation in pregnancy using healthcare databases, the approach being extended from point treatments to treatment changes.","marker":"[40]"},{"why":"Supplies sensitivity and quantitative bias analysis methods the paper recommends alongside these designs.","marker":"[49]"}],"fun_headline_variants":["Start pregnancy drug trials at the clinic, not conception","For pregnancy meds, start the study at first prenatal visit","Avoid immortal time bias: anchor drug studies at prenatal care","Rethink time zero: pregnancy drug trials start at prenatal visits"],"cache_read_input_tokens":13952,"weakest_assumption_plain":"The whole design rests on prescription fills within the grace period correctly revealing which treatment strategy a person actually follows; if fills misrepresent true use through stockpiling, leftover medication, provider samples, or non-adherence, the clone censoring and weights assign people to the wrong strategy and the estimates are biased.","fun_headline_variants_meta":{"raw":{"variants":["Start pregnancy drug trials at the clinic, not conception","For pregnancy meds, start the study at first prenatal visit","Avoid immortal time bias: anchor drug studies at prenatal care","Rethink time zero: pregnancy drug trials start at prenatal visits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":3066,"prompt_tokens":909,"completion_tokens":2157,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2088}},"tokens_in":525,"tokens_out":2157,"duration_ms":13249,"temperature":1.0,"reasoning_tokens":2088,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:02:17.855488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A head-to-head emulation of the same treatment-change question in the same administrative data, run once with conception as time zero and once with the first prenatal encounter as time zero, would settle the claim if the two estimates diverge in the direction and magnitude predicted by immortal time bias; if they do not diverge, re-anchoring time zero changes nothing.","supporting_citations":[],"review_version":1}