Target trial emulation is a methodological framework that analyzes observational data by first specifying, in full protocol detail, the hypothetical randomized controlled trial — the “target trial” — that would answer the research question, then emulating that protocol in the analysis. Formalized by Hernán and Robins in a 2016 American Journal of Epidemiology paper.
The core insight is procedural, not statistical: many of the systematic errors that make observational comparative-effectiveness studies unreliable — immortal time bias, prevalent-user bias, confounding by indication — arise not from any single flawed statistical technique but from the analysis never having a clearly specified trial protocol to be faithful to in the first place. Writing that protocol down explicitly, before touching the data, forces design decisions (eligibility, treatment assignment, start of follow-up, outcome definition) that observational studies frequently leave implicit or get wrong by default.
The problem target trial emulation was designed to solve
Observational studies of treatment effects — using claims data, electronic health records, or disease registries — are attractive because they’re often larger, cheaper, faster, and more representative of real-world practice than a randomized controlled trial (RCT). But without the structural protections randomization provides, several well-documented biases recur across the observational comparative-effectiveness literature:
- Immortal time bias — a period of follow-up during which, by construction of the study design, a participant cannot experience the outcome (often because treatment status is defined using information only available after baseline) gets misclassified in a way that spuriously favors the treated group. This is one of the most common and best-documented design errors in observational pharmacoepidemiology.
- Prevalent-user bias — enrolling patients who are already established on a treatment, rather than new (incident) users, mixes early treatment effects and early dropout patterns into the comparison in ways a trial’s randomization-at-baseline structure avoids.
- Confounding by indication — when the reason a clinician chose a treatment for a given patient is itself related to that patient’s prognosis, comparing treated versus untreated patients conflates the treatment’s effect with the reason it was prescribed.
- Ill-defined or shifting eligibility and treatment-assignment windows — without an explicit protocol, it’s easy for eligibility criteria, the definition of “time zero” (when follow-up starts), and the treatment strategies being compared to be defined inconsistently or discovered post hoc from the data, which can introduce bias even when the underlying data are otherwise sound.
Hernán and colleagues have argued, including in a related 2016 Journal of Clinical Epidemiology paper focused specifically on immortal time bias, that explicitly specifying a target trial protocol before analysis prevents many of these errors by construction — because a design decision that would be nonsensical in an actual randomized trial (like defining “time zero” using data collected after enrollment) becomes visibly nonsensical once the target trial’s protocol is written down for comparison.
The seven-component protocol framework
The original Hernán & Robins framework specifies the target trial’s protocol along the same structural components a real trial protocol would need, and then requires the observational emulation to specify how each one is emulated using the available real-world data:
- Eligibility criteria — who would have qualified for the target trial, defined in a way that can also be applied consistently to the observational data.
- Treatment strategies — the specific interventions being compared (e.g., “initiate drug A” vs. “initiate drug B,” or a specific treatment algorithm), not vaguely defined exposure categories.
- Assignment procedures — in the target trial, participants would be randomly assigned; in the emulation, since randomization isn’t possible, this component describes the statistical strategy (e.g., adjustment for measured confounders, propensity-score methods, or instrumental-variable approaches) used to approximate the balance randomization would have produced.
- Follow-up period — when follow-up starts (“time zero”) and ends, defined so that time zero coincides with both eligibility and treatment assignment — the design decision most directly responsible for preventing immortal time bias when specified correctly.
- Outcome definition — the outcome(s) of interest, defined precisely enough to be measured consistently in the real-world data source.
- Causal contrast of interest — typically the intention-to-treat effect (comparing groups as assigned/initiated, regardless of subsequent adherence) and/or the per-protocol effect (comparing groups as if adherence to the assigned strategy had been maintained), each of which requires a different emulation strategy.
- Analysis plan — the statistical approach used to estimate the causal contrast, pre-specified rather than chosen after seeing results.
The discipline of the framework is in requiring every one of these seven components to be specified explicitly for the target trial first, and only then asking, component by component, how each is emulated (or where the emulation necessarily falls short) using the observational data source actually available. Where a component can’t be adequately emulated — for example, if the data source can’t reliably capture the intervention as randomized in the target trial — that’s a signal the emulation itself may not be trustworthy for that question, which is a diagnostic the framework is specifically designed to surface rather than obscure.
What target trial emulation does and doesn’t fix
Target trial emulation is a design discipline, not a statistical estimator, and it does not solve every limitation of observational data. Specifically:
- It substantially reduces self-inflicted, design-induced biases — the ones that arise from how the study is structured (immortal time bias, prevalent-user bias, inconsistent time-zero definitions) — because those are exactly the errors that become visible once compared against an explicit trial protocol.
- It does not, on its own, solve unmeasured confounding — if a factor that influences both treatment choice and outcome was never recorded in the underlying data source, no amount of careful protocol specification recovers it. The framework makes the assumptions required for a valid causal estimate explicit (and testable/assessable, where possible) rather than eliminating the need for them.
- It depends entirely on the underlying data source’s fitness for purpose — claims data, EHR data, and registries each have their own measurement limitations (e.g., claims data often lack lab values or indication-for-use detail EHR data may capture, and EHR data can have missingness or documentation-practice variation claims data don’t), and the emulation is only as good as the data’s ability to instantiate each of the seven protocol components.
Because of this, target trial emulation studies routinely report their causal assumptions and data limitations alongside their results, rather than presenting the emulated estimate as automatically equivalent to a real trial’s.
Target trial emulation vs. digital twins / synthetic control arms
Target trial emulation is sometimes discussed alongside AI-generated digital twins and synthetic control arms, but the two are distinct methodological tools solving different problems:
- Target trial emulation is a protocol-specification discipline applied to existing, already-collected observational data (claims, EHR, registries) to estimate a treatment effect as if a randomized trial had been run, when no such trial exists or is feasible. Its core mechanism is rigorous causal-inference protocol design applied retrospectively (or prospectively, to data as it accrues).
- Digital twins / synthetic control arms, by contrast, are used within a prospective, ongoing trial to generate individualized, model-predicted outcomes for enrolled participants under a reference condition, in order to reduce the size of the real control arm needed while preserving statistical power. The mechanism is predictive modeling of an individual’s counterfactual trajectory, not protocol emulation against a hypothetical trial.
In short: target trial emulation asks “how would this hypothetical trial’s answer look if we carefully rebuild it from real-world data we already have,” while digital twins ask “can a model’s prediction stand in for part of a real control arm in a trial we are actively running.” They can, in principle, both draw on real-world data as an input, but they are not interchangeable frameworks and shouldn’t be cited as if one were a special case of the other.
Where target trial emulation is used
Common applications include comparative-effectiveness research using claims or EHR data, pharmacoepidemiology safety and effectiveness studies conducted after a drug is already on the market, and generating supportive real-world evidence in regulatory contexts where a confirmatory RCT is not available or not ethical (for example, in some rare-disease or oncology settings). It is also increasingly used as a teaching framework in epidemiology methods training specifically because it makes the causal-inference reasoning behind an observational study’s design explicit and checkable, rather than implicit.
Regulators and health technology assessment bodies evaluating real-world evidence (RWE) submissions increasingly expect sponsors to be able to articulate their observational study design in target-trial terms — explicit eligibility, treatment strategies, time zero, and analysis plan — as a baseline for assessing whether the observational analysis is fit to support the claim being made, though the specific regulatory expectations vary by agency and jurisdiction and should be confirmed against current guidance for any given submission.
Frequently asked questions
Who developed target trial emulation?
The framework was formalized by Miguel A. Hernán and James M. Robins in “Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available,” published in the American Journal of Epidemiology in 2016 (Hernán MA, Robins JM. Am J Epidemiol. 2016;183(8):758–764). A related 2016 paper by Hernán, Sauer, Hernández-Díaz, Platt, and Shrier in the Journal of Clinical Epidemiology focused specifically on how the framework prevents immortal time bias. The framework builds on decades of earlier causal-inference methodology, but this pairing of papers is the standard citation for the formalized approach.
Is target trial emulation the same as a pragmatic trial?
No. A pragmatic trial is a real, prospective randomized trial designed to reflect routine clinical practice conditions as closely as possible while still randomizing participants. Target trial emulation involves no randomization at all — it is a design discipline for analyzing observational data as if a trial (real or hypothetical) had been run. The two can be complementary: a target trial emulation is sometimes explicitly benchmarked against a real RCT’s results (when one exists) to validate the emulation approach for a given data source and clinical question.
Does target trial emulation eliminate confounding?
No. It reduces design-induced, self-inflicted biases that arise from how an observational study is structured, but it cannot recover information a data source never captured. Unmeasured confounding — a variable that affects both treatment choice and outcome but was never recorded — remains a limitation that depends on the richness of the underlying data source and the analytic methods (e.g., adjustment, propensity scores, sensitivity analyses) used within the emulation’s assignment-procedure component.
What is immortal time bias, briefly?
Immortal time bias occurs when a stretch of follow-up time is, by the study’s design, a period during which a participant cannot experience the outcome being studied — typically because treatment-group assignment is determined using information that becomes available only after the start of follow-up. This misclassified “immortal” period spuriously inflates the apparent benefit of the treatment being studied. Explicitly defining time zero as the point of both eligibility and treatment assignment, as the target trial framework requires, is one of the main structural protections against it.
What data sources are used for target trial emulation?
Most commonly administrative/insurance claims data, electronic health record (EHR) data, and clinical or disease registries. Each source has its own strengths and measurement gaps, and the choice of data source directly constrains how faithfully each of the seven protocol components can actually be emulated for a given research question.
