Target Trials and Protocol-First Causal Design

979 people in Saskatchewan, all aged 55 or over, all hospitalised for the first time with chronic obstructive pulmonary disease between 1990 and 1997. Each was followed for one year from discharge. In that year 389 of them were readmitted or died. Samy Suissa did not have to invent a cautionary study. In 2003 he took this one apart.

One cohort, analysed two ways. The first way was the way the literature was doing it. Anyone with a dispensing of an inhaled corticosteroid within 90 days after discharge counts as treated, everyone else as untreated. The clock starts at discharge for both. The second way let a person's exposure status change on the day the dispensing actually happened. The abstract gives both answers in one sentence: “The time-fixed adjusted rate ratio was 0.69 (95% CI: 0.55–0.86) for inhaled corticosteroid use within 90 days, whereas the time-dependent rate ratio was 1.00 (95% CI: 0.79–1.26).” Same 979 people. Same 389 events. Same covariates. A protective effect in the first analysis; nothing at all in the second. The title of the paper names the culprit: immortal time bias.

Look at what the first definition quietly demands of a treated person. To be counted as treated they have to still be alive, still out of hospital, on the day the prescription is filled. That day may be the day after discharge, or as much as 90 days later. An untreated person is under no such obligation. They can have the outcome that same afternoon.

So the treated group arrives carrying a stretch of time in which, by construction, nothing bad is permitted to occur. The drug did not buy that time. The definition did. Widen the exposure window from 15 days to 365 days and the time-fixed rate ratio slides from 0.98 down to 0.51. On the very same data, the time-dependent estimate stayed between 1.06 and 0.94. The apparent benefit grew with the length of the waiting period, not with any dose of medicine.

The repair is not a better model, and no amount of adjustment reaches it. The repair is a shared baseline: one moment at which everyone is checked, assigned a plan, and started on the clock together. It is the moment a randomized trial would have used. This design never had one.

A target trial is the randomized experiment that would settle the question, written out in full before anyone looks at what the data contains. Its parts are the parts of any trial protocol. Who is eligible, which treatment strategies are compared, how people are assigned to them, when the clock starts. Then how long they are followed, which outcome is recorded, which contrast is being estimated, and how the analysis will produce it.

None of that requires data. That is the whole point of writing it first. The protocol fixes the question, and the data is then asked whether it can answer that question — instead of being allowed to decide what the question was.

This is no longer a preference expressed in methods journals. It is in regulatory text. ICH's harmonised guideline M14 was adopted in September 2025 and taken up by the European Medicines Agency the same month. The FDA issued it as final guidance in March 2026. It comes into effect on 18 March 2026. On the research question, the guideline says plainly: “When the study has a causal (inferential) objective, an approach such as the target trial emulation may be used.” Among its references is Hernán and Robins, 2016.

Emulation is the second step: taking each line of the protocol and pointing at the recorded field that will stand in for it. The exercise earns its keep mostly where the pointing fails. That is where immortal time surfaces, along with treatment start dates nobody can pin down, adherence that was never captured, and outcomes that were looked for harder in one group than in the other.

The strongest evidence that the protocol was the thing going wrong, rather than the dataset, came from re-opening a famous disagreement. Decades of observational evidence had suggested hormone therapy reduced coronary heart disease risk. Then the Women's Health Initiative randomised 16,608 women and followed them a mean of 5.2 years. It reported the opposite direction: a hazard ratio of 1.29 (nominal 95% CI 1.02–1.63, 286 cases). In 2008 Hernán and colleagues re-analysed the observational Nurses' Health Study as an explicit emulation of that trial. The same cohort now returned intention-to-treat hazard ratios of 1.42 (95% CI 0.92–2.20) for the first 2 years and 0.96 (0.78–1.18) over the entire follow-up. Their conclusion: “Our findings suggest that the discrepancies between the Women's Health Initiative and Nurses' Health Study ITT estimates could be largely explained by differences in the distribution of time since menopause and length of follow-up.” The rows had not changed. The protocol had.

Set three things side by side and the distinction sharpens. A target trial states the experiment the question demands. An emulation maps each component of that experiment onto records that already exist. A database study skips both and groups rows by what people turned out to do later. That is how the convenient columns end up choosing the science.

Take the components in the order a real trial would apply them. Eligibility comes first and decides who is in the study at all. Strategies come next — the specific plans being compared, each one something a person could be handed rather than a label earned by later behaviour. Assignment puts each eligible person onto one of those strategies. Follow-up runs from that moment until the outcome or until censoring. The estimand and the analysis close the sequence: which contrast is being estimated, and by what procedure.

Three of those lines have to land on the same instant — eligibility judged, strategy assigned, follow-up begun. That instant is time zero. Hernán and four colleagues set the alignment principle out formally in 2016. Specifying the target trial first, they argue, is what prevents immortal time bias and other self-inflicted injuries in observational analyses. They catalogue four distinct ways an emulation breaks the alignment.

Their fourth failure is the one worth memorising, because it looks like diligence. Assign the strategies using treatment observed after time zero — say, requiring three filled prescriptions before a person counts as a user — and the label is no longer something anyone could have been handed at baseline. Emulation failure 4 puts it this way: “Anyone who filled a third prescription of aspirin is retrospectively declared to have been immortal during the first year of follow-up.” The requirement was meant to identify real users. What it actually identified was people who survived long enough to become them.

It is also, exactly, where Suissa's 90-day window came apart. Reading the protocol in order is how you find that out before publishing rather than after.

Once the protocol is on paper, the familiar catalogue of observational errors stops reading like a list of separate traps. It starts reading like a single thing said four ways: a line of the protocol that the data was quietly allowed to violate. The regulators reached the same conclusion and wrote it into the guideline. On time-related bias, ICH M14 prescribes the design rather than a correction: “These risks may be mitigated by design frameworks (see Section 4.1, Research Question), as this approach aligns assessment of eligibility and baseline information with start of follow-up.” Each of the following names one line.

Pick one decision you already believe something about, and write its ideal trial before touching a table. Who would have been eligible, on what day, judged on what information. Which strategies you would have compared. How the clock starts, and for whom it starts when.

Then open the data with that page beside you and work down it. Of each line, ask not whether some field exists but whether that field means what the protocol needs it to mean. Where the exposure is defined by a window, run Suissa's test on your own definition before anyone else does. Move the window and see whether the answer moves with it. His did — 0.98 at 15 days, 0.51 at 365 — while the estimate that respected the clock stayed between 1.06 and 0.94.

Keep the lines you could not fill. That list is your limitations section, written before the analysis instead of negotiated after it.

Four terms carry most of the weight above, and two of them are really about a clock. They are easy to blur together and expensive to blur.

An emulation deserves a causal reading under five conditions. One instant has to serve as eligibility, assignment and the start of follow-up for everyone. The strategies compared have to be ones a person could genuinely have been handed at that instant. Measured confounding and censoring have to be addressed by a stated method rather than a hope. Every deviation from the ideal trial has to be written down. And the report has to say which population the estimate actually describes.

That last requirement now has a form to fill in. The TARGET statement, published in JAMA in 2025, is a 21-item checklist across 6 sections. Behind it sit a systematic review, a two-round survey of 18 experts from 6 countries between August 2023 and March 2024, and a three-day consensus meeting of 18 panelists in June 2024. It was then piloted with 108 stakeholders between September 2024 and February 2025. Journals have begun to require it. On 31 October 2025 PLOS Medicine put it plainly: “PLOS Medicine is announcing that all manuscripts submitted to the journal that rely on TTE (or claim to do so) must adhere to the TARGET reporting guideline as a submission requirement”.

Do all of it and the conclusion is still observational. Nothing in the procedure randomized anybody. The 32 pre-registered emulations correlated with their trials at r = 0.82 overall, and at r = 0.53 wherever the records could not carry the protocol.

What the discipline buys is a claim whose weaknesses can be listed. A reader can see the assumptions, see the gaps left by the records, and disagree with one specific line rather than with a mood. That is worth a great deal. It is not the same thing as evidence from a trial, and borrowing the word does not borrow the design.