What is target trial emulation?
What target trial emulation is, its seven protocol parts, why time zero decides the design, the effects it estimates, what it cannot fix and how to report it.
Target trial emulation is a way of designing and analysing an observational study so that it asks the question a randomised trial would have asked. You first write down the protocol of the trial you would run if you could, the target trial, and then use the observational data to emulate each part of it. It is a framework for causal inference, not a statistical method, and it does not remove confounding by itself.
The framework is set out by Hernán and Robins (2016). Its main value is in the design. Many avoidable biases in comparisons of treatments, such as immortal time and the selection of people who are still taking a drug, are failures to copy the basic structure of a trial, and writing the protocol first makes them visible (Hernán et al., 2016).
Why emulate a trial
Observational studies of treatments have a history of disagreeing with trials, and hormone therapy and coronary heart disease is a well-known example. Earlier observational studies had suggested a lower risk of coronary heart disease in women taking oestrogen plus progestin, and the Women's Health Initiative (WHI) randomised trial found a higher one. Hernán et al. (2008) analysed the Nurses' Health Study as a sequence of trials, in each of which eligible women were classified as initiators or non-initiators of the treatment. Their intention-to-treat hazard ratios for initiators against non-initiators were 1.42 (95% confidence interval 0.92 to 2.20) over the first two years, against 1.68 (1.15 to 2.45) in the WHI, and 1.00 (0.78 to 1.28) over the first eight years, against about 1.24 (0.97 to 1.60). Much of the remaining difference disappeared when the authors stratified by time since menopause at the start of therapy: in women within ten years of menopause the hazard ratio was 0.84 (0.61 to 1.14) against 0.88 (0.54 to 1.43) in the WHI.
How far this can be trusted has been tested in one large initiative. RCT-DUPLICATE emulated 32 randomised trials of medicines, selected because they were feasible to emulate, with new-user cohort studies in three US claims databases, and registered all 32 protocols on ClinicalTrials.gov before the outcome analyses. The trial and database estimates had a correlation of 0.82 (95% confidence interval 0.64 to 0.91). In a post hoc split it was 0.93 (0.79 to 0.97) for the 16 trials classed as closely emulated and 0.53 (0.00 to 0.83) for the other 16 (Wang et al., 2023).
The protocol: seven parts
The target trial protocol has seven key components, and each has to be emulated with the data. In Hernán and Robins' (2016) example, the effect of postmenopausal hormone therapy on the five-year risk of breast cancer, they are:
- Eligibility criteria. Who would be enrolled. In the data, apply them using only what is known at time zero, never information from the future.
- Treatment strategies. The interventions compared, written as strategies over time, for example starting therapy at baseline and staying on it unless a serious complication such as a thrombosis, a heart attack or cancer occurs, against not taking it. They are not “exposed” and “unexposed” as the data happen to record them.
- Assignment procedures. In the trial, randomisation. In the emulation, people are not randomised, so you compare people who are alike on the baseline prognostic factors that also influence which strategy a person starts. That works only if those factors are measured, the assumption of no unmeasured confounding.
- Follow-up period. It starts at time zero and ends at the outcome, death, loss to follow-up or the end of the planned follow-up.
- Outcome. What is measured, and how it is ascertained.
- Causal contrasts. The effects to estimate: the intention-to-treat effect, the per-protocol effect, or both.
- Analysis plan. The estimator, the confounders and how censoring is handled, written before the outcome results are seen.
Time zero
The part most often done wrongly is time zero. In a trial, eligibility is assessed, a strategy is assigned and follow-up starts at the same moment, at randomisation. In observational data these moments drift apart, and the way they drift creates bias. If follow-up starts at diagnosis for the people who later start treatment but they must survive until they start it, the treated have immortal time (What is immortal time bias?). If the treated are picked up only after they have been on treatment for a while, as prevalent users, the early events and the early stops that new users have are missing from the treated group.
The emulation sets time zero at the moment an eligible person is assigned to a strategy on the basis of the data up to that moment, so that eligibility, assignment and the start of follow-up coincide, as they do at randomisation. One way to do it is the sequence of trials of the 2008 study: a new trial begins at each time point at which a person is eligible, so the same person can appear in several of them, and the variance has to allow for that.
Some strategies are not decided at time zero, for example starting treatment within six months of diagnosis. A usual remedy is to clone each person into every strategy at time zero, to censor each clone when the person stops following that strategy, and to weight the remaining follow-up to correct for the selection that the censoring creates, which works only if the weighting captures the reasons people deviate. Because cloning puts every person into every strategy at time zero, this design estimates a per-protocol effect and cannot give an intention-to-treat effect, and model-based standard errors are too small, so confidence intervals usually come from the bootstrap. Maringe et al. (2020) give a step-by-step tutorial with R and Stata code. In their example of surgery in older patients with early-stage lung cancer, a naive Kaplan–Meier analysis put the benefit to one-year survival at 22 percentage points, the cloned analysis without weights at 17, and the weighted analysis at 11.
Intention-to-treat and per-protocol effects
A trial's primary analysis is often the intention-to-treat effect, the effect of being assigned to a strategy whatever happens next. Its observational counterpart compares people by the strategy they started, and, as in a trial, it tends to be diluted when people stop or switch. The per-protocol effect, the effect of following a strategy throughout, usually requires people to be censored when they deviate and the remaining follow-up to be weighted for the baseline and later factors that predict both deviating and the outcome. Because those later factors can themselves be affected by earlier treatment, g-methods such as inverse probability weighting are needed. The 2008 study reported both: its adherence-adjusted hazard ratios, estimated by inverse probability weighting, were 1.61 (0.97 to 2.66) over the first two years and 0.98 (0.66 to 1.49) over the whole follow-up, with wider intervals than the intention-to-treat estimates.
What it does not fix
- Unmeasured confounding. The emulation can only adjust for what was measured. It makes the assumption explicit and does not make it true.
- Data that lack what the protocol needs. If the data cannot identify who was eligible, when a strategy started or the main confounders, the emulation is partial, and the protocol should say where. In a post hoc comparison in RCT-DUPLICATE, agreement was closest for the trials that could be emulated closely.
- Measurement and design limits. Treatment, confounders and outcomes recorded for other purposes can be misclassified, and only pragmatic, unblinded trials can be emulated.
- Strategies that are not interventions. “Having a lower body mass index” is not a well-defined intervention, because different ways of lowering it can have different effects. The question has to be one that a trial could in principle ask.
Reporting an emulation
The TARGET guideline, published in 2025, is a 21-item checklist for reporting observational studies that emulate a target trial (Cashin et al., 2025). Its key recommendations are to identify the study as an emulation of a target trial, to summarise the causal question, to specify the target trial protocol and how it was mapped to the data, and to report the estimate for each estimand with its precision and with sensitivity analyses of the assumptions and of the design and analysis choices. It covers parallel-group, individually randomised target trials with adjustment for baseline confounders.
Designing an emulation?
The protocol, the choice of time zero and the handling of strategies that are not fixed at baseline are where an emulation stands or falls, and where a reviewer or a regulator will look first. We design and analyse them for register and other real-world data. See our real-world evidence services, or talk to us about your question.