Sequential Target Trial Emulation

In plain language

Sequential target trial emulation turns a longitudinal database into a series of small emulated trials. A patient can be eligible at several decision times, such as each month or each clinic visit, and each decision time gets its own time zero. This is useful when treatment can start at many opportunities, but the pooled result is an average over those opportunities, not necessarily the effect for one fixed baseline population.

Sequential target trial emulation

extends target trial emulation from one baseline date to many repeated baseline dates. Instead of asking only "what happened to patients eligible on January 1?", the analyst defines a trial protocol that could begin at each month, visit, prescription opportunity, or risk-set time. At each trial start, eligible patients are cloned into a person-trial, assigned according to the treatment strategy observed at that start, followed from that local time zero, and censored when they deviate from the assigned strategy if the estimand is per-protocol. The person- trials are then pooled with trial-start indicators and robust variance or clustered inference.

Core conceptual distinction

A single target trial has one eligibility/assignment/follow-up origin. A sequential emulation has many origins, each with its own time zero. This is useful when treatment can be initiated at many clinical opportunities and non-initiators at one opportunity may legitimately remain eligible for later opportunities. The unit of analysis is therefore not simply the patient; it is the patient-at-trial-start. The same patient may appear as an eligible non-initiator in January, again in February, and then as an initiator in March, depending on the protocol.

The estimand must be named carefully. A pooled sequential emulation usually estimates an average effect over the set of eligible treatment opportunities represented by the trial starts. That is not automatically the same target population as a one-time trial at diagnosis or first eligibility. If treatment effects vary by calendar time, disease duration, severity, variant era, or time since diagnosis, pooling across trial starts estimates a weighted mixture of opportunity-specific effects. Trial-start heterogeneity is therefore a design feature to report, not a nuisance to hide.

Pros, cons, and trade-offs

- vs a single-baseline target trial emulation: Sequential emulation uses more of the longitudinal data and avoids discarding patients who were untreated but still eligible after the first opportunity. Cost: the target population is less intuitive, rows are correlated within patient, and pooled effects can mask trial-start heterogeneity. Prefer a single-baseline emulation when the clinical decision is truly one-time; prefer sequential emulation when the decision recurs. - vs marginal structural models fit directly to person-period data: Both can target time-varying treatment strategies. Sequential emulation is often easier to explain because every row maps to a target-trial protocol component. Cost: data expansion can be large, and artificial censoring plus inverse-probability weights still require the same time-varying confounder assumptions. Prefer sequential emulation for protocol transparency and auditability. - vs naive time-dependent exposure Cox models: A time-dependent model may avoid immortal time in a narrow technical sense, but it often answers a conditional association among survivors rather than the effect of initiating at a decision point. Sequential emulation anchors each decision at its own time zero and makes eligibility, assignment, and follow-up explicit. - vs clone-censor-weight for one sustained strategy: Clone-censor-weighting can be implemented inside each sequential trial when strategies are sustained or dynamic. Sequential emulation supplies the repeated decision origins; cloning and weighting supply the per-protocol machinery.

When to use

Use sequential emulation when the treatment decision repeats and the data contain meaningful repeated eligibility assessments: monthly initiation of a chronic medication, vaccination during an outbreak, biologic escalation at rheumatology visits, dialysis modality switching, oncology maintenance therapy after cycles, or guideline-directed treatment intensification after repeated lab thresholds. It is also useful when a strict first-eligibility cohort would be too small or would answer an artificial one-time question.

When NOT to use - and when it is actively misleading

- Do not use it to manufacture precision when the repeated trials do not share a defensible estimand. If eligibility in later trials represents a much sicker or more treatment-resistant survivor population, the pooled estimate may not answer a policy-relevant question. - Do not let the same future information define both eligibility and assignment. Eligibility, treatment status, covariates, and follow-up must be evaluated at the trial start using information available at that time. - Do not pool blindly across calendar periods with strong effect modification, such as changing viral variants, product supply, label restrictions, or clinical guidelines. Report trial-start-specific estimates or interactions. - Do not censor non-initiators at later initiation without inverse-probability-of-censoring weights when estimating a sustained "do not initiate" or "remain untreated" strategy. That censoring is typically informative. - Do not interpret row counts as patient counts. A sequential dataset may have 10 person-trials for one patient and one for another; summaries must report both.

Data-source operational depth

- Claims: Build a discrete trial calendar, commonly monthly. At each candidate start, require observable enrollment and a clean lookback for eligibility and confounders. Assignment is usually first qualifying fill during a grace window after the trial start versus no initiation. Censor at disenrollment, death, outcome, or strategy deviation according to the estimand. Because one person can contribute multiple person-trials, use patient-clustered robust standard errors, bootstrap by patient, or influence-function methods that respect within-person correlation. - EHR: Trial starts can be visits or lab-threshold dates rather than calendar months. This improves clinical relevance but creates irregular opportunities: patients with more visits have more chances to enter trials. Define visit processes, eligibility recency windows, and missing-lab rules explicitly so visit intensity does not become hidden selection. - Registry: Registries with scheduled follow-up visits are well suited because eligibility can be assessed uniformly. The danger is late registry entry: if the registry enrolls prevalent treated patients, early opportunity trials are missing. Link to claims or EHR to reconstruct pre-registry treatment history where possible. - Linked data: Linked claims-EHR-registry data allow richer time-varying confounding control but amplify date reconciliation problems. Decide whether time zero is visit date, lab date, fill date, or administration date before data expansion.

Worked example

A diabetes study asks whether starting an SGLT2 inhibitor versus not starting one reduces 2-year heart failure hospitalization among adults already using metformin. Patients become eligible in any month when they have type 2 diabetes, continuous enrollment, no SGLT2 inhibitor fill in the prior year, no prior heart failure hospitalization, and at least one recent HbA1c measurement. In January, patient 410 is eligible and does not initiate, so they enter the January trial as a non-initiator. In February they are still eligible and again do not initiate, so they enter the February trial. In March they fill an SGLT2 inhibitor within the protocol grace period, so they enter the March trial as an initiator. Follow-up for each person-trial starts at that trial's month-specific time zero. For an ITT-like initiation contrast, the March trial follows patient 410 as an initiator regardless of later switching. For a per-protocol contrast, the January and February non-initiator clones are censored when the March initiation violates their assigned strategy, and weights are used to handle that informative censoring.

Worked example

Scenario

Patient 410 is eligible to start an SGLT2 inhibitor in January, February, and March. They do not start in January or February, then initiate in March. A sequential emulation represents those repeated decisions as separate person-trials.

Dataset

Person-trial rows created from one patient's monthly eligibility history.

Steps

• Define the target trial once, including eligibility, strategies, time zero, follow-up, outcome, and analysis.

• Repeat that protocol at each monthly trial start.

• Stack eligible patient-months into person-trials, preserving the local time zero for each row.

• For an ITT-like contrast, follow each row according to assignment at the local trial start.

• For a per-protocol contrast, censor rows when their observed data stop following the assigned strategy and use IPCW.

Result

Patient 410 contributes two non-initiation person-trials and one initiation person-trial. The analysis must cluster by patient and state that the pooled effect averages over eligible monthly treatment opportunities.