Wait, what? One of the best ways to analyse observational data is to begin by designing a randomised trial that never happened.

That sounds backwards, but it solves a deep problem. Many observational analyses begin with a convenient dataset and ask what associations can be extracted from it. Target trial emulation reverses the order: first define the causal experiment that would answer the question, then ask how closely the available observational data can emulate that experiment.

The one-sentence answer

Target trial emulation is a causal-design framework in which researchers explicitly specify the protocol of the hypothetical randomised trial they would ideally run, then map observational data onto that protocol as faithfully and transparently as possible.

Quick Read

• Start with the causal question, not the database.
• Define who would be eligible, which strategies are compared, when treatment assignment occurs, when follow-up starts, which outcomes matter and which causal contrast is wanted.
• Align eligibility, treatment assignment and start of follow-up at a common time zero.
• Observational data cannot reproduce randomisation; measured confounding still has to be addressed.
• The framework helps expose immortal-time bias, prevalent-user bias, selection problems and ambiguous treatment definitions.
• Good emulation depends on data quality, temporal resolution, treatment measurement, outcome definition and overlap.
• Transparent reporting matters because two studies can use the same database and emulate different target trials.

Why “use the data we have” is not enough

Suppose a school system wants to know whether beginning an intensive revision programme in January improves examination outcomes compared with beginning in March. A database contains attendance, programme participation and final grades. It is tempting to compare January participants with March participants.

But when exactly did each student become eligible? Were students already participating before records began? Did the January group have to remain enrolled long enough to be classified as “January starters”? Were students assigned according to prior performance? Was follow-up counted from the same moment for both strategies? Did some students switch strategies? What effect is being estimated: assignment to a strategy, adherence to it, or actual time exposed?

These are design questions, not merely statistical questions. If they are left vague, no advanced model can guarantee that the resulting comparison answers a coherent causal question.

The target trial protocol

A target trial makes the intended causal experiment explicit. The exact terminology varies slightly across sources, but the protocol normally specifies several core components.

• Eligibility criteria: who could enter the trial?
• Treatment strategies: what precisely are the alternatives being compared?
• Assignment procedure: how would participants be assigned in the ideal trial?
• Time zero: when do eligibility, assignment and follow-up begin?
• Outcomes: what will be measured, over what horizon and with what definitions?
• Causal contrast or estimand: which effect is the target—such as assignment to a strategy versus sustained adherence?
• Analysis plan: how will the target effect be estimated?

The observational emulation then documents how each protocol element is represented in the available data and where the representation is imperfect.

Time zero: the alignment that prevents many injuries

One of the framework’s most important disciplines is aligning three events: meeting eligibility, assigning the treatment strategy and beginning outcome follow-up.

In a randomised trial, these events are deliberately coordinated. In observational data they can drift apart. That drift can create serious bias.

Consider a study classifying students as “completed six months of tutoring” and then comparing them with non-participants from the beginning of the school year. To enter the six-month group, a student must remain observable and tutoring long enough to qualify. The period before classification is effectively guaranteed survival in the treatment group. In health research this structure is famously associated with immortal-time bias; the broader logic applies whenever future information is used to define a baseline group.

The target-trial question forces a correction: what would assignment have looked like at the actual beginning?

New-user design: do not let survivors define the treatment

Another common observational problem is comparing people who are already long-term users of an intervention with people who are not. Long-term users have, by definition, survived the early period, tolerated the intervention and remained in the relevant system. They can therefore be a selected group.

A target trial often motivates a new-user design: compare eligible people when they initiate one strategy or the alternative. This moves the observational comparison closer to the moment at which an ideal trial would have assigned treatment.

Treatment strategies must be operational, not decorative

“Use tutoring” is not a sufficiently precise treatment strategy if the causal question depends on frequency, start date, duration or switching. A useful strategy might be: begin two 90-minute sessions per week within seven days of eligibility and continue through the examination period, versus begin the same schedule eight weeks later.

The more precisely the strategy is defined, the easier it becomes to determine whether the available data can actually identify adherence, deviation, switching and exposure duration.

Assignment is the piece observational data cannot truly reproduce

Randomisation breaks systematic links between baseline characteristics and assigned treatment, apart from chance. Observational treatment assignment does not. People, institutions, circumstances and prior states influence what happens.

Target trial emulation therefore does not “make” the data randomised. It specifies the random assignment procedure the ideal trial would use, then replaces randomisation analytically by an assumption: conditional on the measured covariates used for adjustment, treatment assignment is sufficiently unconfounded for the causal contrast being estimated.

That assumption must be defended, not hidden.

Baseline confounding: design before estimation

Regression adjustment, matching, stratification, propensity-score weighting and doubly robust estimators can all play roles in the emulation. But the target-trial protocol comes first. Otherwise the analyst risks choosing a statistical method without knowing which causal experiment the method is supposed to emulate.

This is a major conceptual advantage of the framework: it separates design from estimation. The design tells us what question is being asked and how the observational records correspond to trial participants and strategies. The estimator then attempts to recover the specified causal contrast under the stated assumptions.

Intention-to-treat versus per-protocol thinking

A randomised trial can estimate the effect of assignment to a strategy even when some participants do not adhere. This intention-to-treat contrast benefits from randomisation because assignment itself is random.

Observational data rarely contains literal random assignment. Researchers may instead estimate an analogue of assignment at initiation or a per-protocol effect of sustained adherence. The latter becomes more complicated because adherence over time can depend on evolving covariates that are themselves affected by earlier treatment.

This is where g-methods, inverse probability weighting for treatment and censoring, cloning-censoring-weighting designs and other longitudinal causal tools can become necessary.

Time-varying confounding is a different problem

Suppose students can change tutoring intensity every month, and monthly performance affects both future tutoring decisions and later examination results. Monthly performance is now both a predictor of future treatment and part of the pathway affected by past treatment.

Ordinary regression adjustment can fail in this setting because controlling for a time-varying covariate affected by earlier treatment can block part of the causal pathway or induce additional bias. Longitudinal g-methods were developed for precisely these structures.

Censoring is not merely missingness

Participants may leave the data source, switch strategies, stop being observed or deviate from the protocol. If those processes are related to prognosis, simply ignoring censored observations can distort the result. An emulation therefore has to define censoring rules and, where appropriate, adjust for informative censoring.

The key question is always causal: what selection process did the analysis create, and what assumptions are required to compare the remaining observations?

Data architecture decides what can be emulated

A perfectly written target-trial protocol does not guarantee that a database can support it. The data must record the relevant eligibility variables before treatment, treatment initiation with adequate timing, confounders at the necessary resolution, outcomes with defensible definitions and enough follow-up to estimate the target contrast.

If treatment is recorded only once per year, a question requiring day-level initiation may not be emulable. If prior exposure history is absent, a clean new-user design may not be possible. If the outcome captures only a proxy, the estimand changes. This is not a software problem. It is an evidence-boundary problem.

Benchmarking against randomised trials

When a genuine randomised trial exists, researchers can sometimes emulate that trial using observational data and compare the estimates. This is a powerful stress test because the target protocol is already known.

A 2026 systematic review and meta-analysis in The BMJ examined 107 pairs of target trial emulations and corresponding randomised trials. Overall concordance was moderate, and closer emulation of trial design was associated with substantially better agreement. That does not establish that emulations are interchangeable with trials; it shows why design fidelity, data quality and transparent benchmarking matter.

Reporting has now become part of the method

In 2025, the TARGET Statement introduced a 21-item reporting checklist for observational studies explicitly emulating a target trial. The guidance emphasises stating the causal question, explaining why emulation is needed, specifying the target trial protocol, mapping that protocol to the observational data, defining the estimand and assumptions, and reporting sensitivity analyses and design choices transparently.

This matters because “target trial emulation” should not become a fashionable label attached to ordinary observational analyses. The protocol must be visible enough for another researcher to understand what was emulated and where the emulation departed from the ideal.

A worked education example

Question: does beginning structured weekly retrieval practice at the start of Term 2 improve end-of-year examination performance compared with beginning the same programme at the start of Term 3?

• Eligibility: students enrolled at the end of Term 1 with complete baseline assessment data and no prior exposure to the programme.
• Strategies: initiate the defined retrieval-practice programme in week 1 of Term 2 versus week 1 of Term 3.
• Time zero: first day of Term 2 for eligibility, strategy assignment and follow-up.
• Outcome: standardised end-of-year assessment score.
• Baseline confounders: prior achievement, attendance, school, subject level and other pre-specified determinants of both programme uptake and outcome.
• Estimand: the average effect of assignment to early versus later initiation in the eligible population.

The observational emulation then asks whether the data can identify all of those elements. If students who “begin in Term 2” are defined using information collected weeks later, time zero is broken. If prior programme exposure is unknown, new-user status is uncertain. If schools adopted the programme selectively, school-level confounding must be addressed. The target trial does not make those problems vanish; it makes them difficult to ignore.

Common failure modes

• Database-first causality: letting available columns define the causal question after the fact.
• Broken time zero: eligibility, treatment assignment and follow-up begin at different times.
• Future-defined groups: using later information to decide baseline treatment membership.
• Vague treatment: calling a heterogeneous range of exposures one intervention without defining a strategy.
• Prevalent-user bias: comparing established users with non-users without accounting for survival and selection into long-term use.
• Confounding amnesia: treating explicit emulation as though it had created randomisation.
• Unsupported adherence effect: estimating sustained-strategy effects without handling time-varying confounding and informative censoring.
• Protocol theatre: writing a target-trial table but making the actual analysis inconsistent with it.

How this connects across the eduKate Library

Target trial emulation is a design bridge rather than a replacement owner. Read it alongside How Causal Inference Works, How Observational Studies Work, How Longitudinal and Panel Data Work, How Propensity Scores, Matching and Inverse Probability Weighting Work, and How Sensitivity Analysis and Robustness Checks Work.

Authoritative and current source corridor

• Hernán & Robins (2016), Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available.
• BMJ guide: Target trial emulation—applying principles of randomised trials to observational studies.
• TARGET Statement (2025): Transparent reporting of observational studies emulating a target trial.
• BMJ 2026 systematic review and meta-analysis of concordance between target-trial emulations and randomised controlled trials.

The deeper idea

Target trial emulation is powerful because it changes the order of thought. Instead of asking, “What can this dataset tell me?”, it asks, “What experiment would answer the question, and what part of that experiment can this dataset honestly reconstruct?”

That shift turns causal analysis from a hunt for impressive coefficients into a discipline of design, time, comparability and explicit assumptions.