Target trial emulation
Why this level?
An observational comparison uses concurrent controls. Statistical adjustment does not change this design level. NHMRC 2009 sourceEmulation removes specific, nameable biases: immortal time, prevalent-user selection, a time zero set at the wrong moment. It does not change what the data can support, because this is still a cohort analysis. That is the point of the framework rather than a limitation of it.
Where the NHMRC 2009 hierarchy names this design, it places it here:
| Research question | Level | When |
|---|---|---|
| Intervention | III-2 | An observational comparison uses concurrent controls. Statistical adjustment does not change this design level. |
What question does this help answer?
It answers a comparative effectiveness question when a trial is impossible, and its real contribution is discipline: it forces every design choice into the open before the analysis begins.
How it works in plain language
You specify the target trial's protocol: eligibility, the treatment strategies being compared, assignment, follow-up start and end, the outcome, the causal contrast, and the analysis plan. Then, for each, you state how the observational data emulate it. Where the emulation fails, the failure is visible rather than buried.
Why it prevents immortal time bias by construction
In a randomized trial, eligibility, assignment and the start of follow-up all coincide at randomization. Writing down that protocol makes any misalignment in the observational analysis obvious, and misalignment of exactly those three is what creates immortal time bias. The approach does not require vigilance; it makes the error structurally visible.
Aligned and misaligned time zero
When not to use it
It is not a way to make weak data strong. Emulation makes the assumptions explicit; it does not make them true, and a target trial protocol drawn on data that cannot support it produces a clearer account of an untrustworthy estimate.
Why Mallard may recommend it
Mallard may recommend framing a retrospective database study as a target trial emulation when the plan describes comparing treatment strategies in routinely collected data.
Key assumptions and design requirements
- Eligibility, treatment assignment and start of follow-up are ALIGNED at a single time zero.
- Treatment strategies are described precisely enough that a clinician could follow them.
- The causal contrast is named: intention to treat, per protocol, or another.
- Confounding control is stated, and the assumption of no unmeasured confounding is made explicit.
- Where the data cannot emulate an element of the protocol, that gap is reported rather than glossed.
Clone, censor, weight
Some strategies cannot be assigned at time zero from the data, for example start treatment within seven days against do not start, because at time zero nobody has yet done either, and assigning people by what they later did recreates immortal time. The clone-censor-weight approach copies each person into every strategy their data are compatible with at time zero, follows each copy, and censors a copy at the moment the person's actual treatment departs from its strategy: the do-not-start copy when treatment begins, or the start-within-seven-days copy at day seven if treatment has not begun. The grace period is the window in which starting still counts as following the strategy, and it is part of the protocol, fixed in advance. Because this censoring depends on prognosis, inverse probability of censoring weights are applied to the copies still followed. The result is a per-protocol effect, and it assumes the variables in the censoring model capture why people did or did not start treatment. The standard error must allow for the same person appearing under more than one strategy, usually by bootstrapping people rather than rows.
What to watch out for
Emulation makes the assumptions explicit; it does not make them true. Writing down the protocol is what exposes eligibility, treatment assignment and follow-up start as things that must coincide, which is why the approach prevents immortal time bias by construction rather than by vigilance.
Clinical example
Question: Does starting anticoagulation within 48 hours reduce stroke? Design: Write the trial you would run, with eligibility at admission, assignment at 48 hours and follow-up starting there, then emulate it in the database with those three aligned rather than defining exposure from anything that happened later.
How to report it
TARGET statement is the guideline for this design. It asks for the emulated protocol in full: eligibility, the treatment strategies compared, how assignment was emulated, when follow-up starts and ends, and the causal contrast, with each element set beside the data actually used to emulate it. A reporting guideline is not a design manual: use it to make the study transparent, not as a substitute for careful design. Check the target journal's own requirements as well, which this page cannot know.
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwv254
- Principles of confounder selection (opens in a new tab) · Eur J Epidemiol · doi 10.1007/s10654-019-00494-6
- How to estimate the effect of treatment duration on survival outcomes using observational data (opens in a new tab) · BMJ · doi 10.1136/bmj.k182
- Reflection on modern methods: trial emulation in the presence of immortal-time bias. Assessing the benefit of major surgery for elderly lung cancer patients using observational data (opens in a new tab) · Int J Epidemiol · doi 10.1093/ije/dyaa057
- Transparent reporting of observational studies emulating a target trial: the TARGET Statement (opens in a new tab) · BMJ · doi 10.1136/bmj-2025-087179