What intention-to-treat analysis actually means

Intention-to-treat, usually shortened to ITT, is the principle that participants in a randomized trial are analyzed in the group they were assigned to, whether or not they actually took the assigned treatment. Its purpose is to preserve randomization, which is the thing that makes the compared groups alike in every respect except assignment.

That definition is accurate. It is also incomplete, and the incompleteness is where most misreading begins. Analyzing people in the group they were randomized to is only one of the principle's consequences, and it is not the demanding one.

What randomization buys, and what ITT protects

Randomization does something no amount of statistical adjustment can fully reproduce. By assigning treatment through a chance mechanism, it distributes both known and unknown differences between participants across the groups. Age, disease severity, comorbidity, motivation, and the many factors nobody thought to measure all tend to balance out.

That balance is what licenses the central inference. If the groups started out comparable and were treated differently, a later difference in outcome can reasonably be attributed to the assignment itself rather than to who ended up where.

Now consider what happens when participants are re-sorted according to what they actually did rather than what they were assigned. This is a per-protocol analysis: it includes only those who adhered to the treatment as specified. It sounds like a fairer test of whether the drug works, and in one narrow sense it is.

The problem is that adherence is not a fixed background characteristic. It is itself an outcome, and it can be influenced by the treatment. People who stop taking a drug frequently stop because of side effects, or because they are getting sicker, or because they feel well enough that they no longer bother. Once the analysis conditions on adherence, the groups are no longer the groups randomization created, and the protection is gone.

Neither analysis is the correct one in the abstract. They answer different questions. The ITT approach estimates the effect of assigning a treatment; a per-protocol approach gestures at the effect of receiving it, but without the safeguard that made the comparison trustworthy in the first place.

Three consequences, not one

The ICH E9(R1) addendum, adopted under Step 4 on 20 November 2019, is unusually clear on this point. It notes that ICH E9 "introduced the Intention-To-Treat (ITT) principle in connection with the effect of a treatment policy in a randomised controlled trial, whereby subjects are followed, assessed and analysed irrespective of their compliance to the planned course of treatment."

From that principle the addendum distinguishes three separate consequences:

• The trial analysis should include all subjects relevant for the research question.
• Subjects should be included in the analysis as randomized.
• Subjects should be followed up and assessed regardless of adherence to the planned course of treatment, and those assessments should be used in the analysis.

The first two are the ones everybody quotes. The third is the one that decides whether a trial can actually deliver on the label.

Consider a plainly illustrative case, not a real trial. Four hundred participants are randomized, two hundred to each arm. Sixty people in the treatment arm stop the drug at week four because of nausea. If the trial keeps measuring those sixty and includes their outcomes, an intention-to-treat analysis is possible. If the trial stops following them the moment they discontinue, their outcomes do not exist, and no analysis method can conjure them back. The methods section may still say the analysis was by intention to treat. It cannot be.

This is why the third consequence is a design commitment rather than an analytic choice. It determines what data get collected, and that decision is made long before anyone opens the dataset.

Intercurrent events, and the shift from a label to a question

The addendum's larger contribution is to reframe the discussion entirely. Rather than asking whether a trial ran an ITT analysis, it asks which treatment effect the trial set out to estimate.

Central to that reframing is the concept of an intercurrent event, which ICH E9(R1) defines as an event "occurring after treatment initiation that affect either the interpretation or the existence of the measurements associated with the clinical question of interest." Stopping the assigned treatment, switching to another, or adding a second medication are the common examples.

The guideline makes a point that is easy to miss and worth sitting with: "Unlike missing data, intercurrent events are not to be thought of as a drawback to be avoided in clinical trials." Discontinuation and the use of additional medication happen in ordinary clinical practice as well as in trials. A trial that treated them purely as contamination would be measuring something further from real care, not closer to it.

The estimand is the precise statement of what is being estimated, specified in advance, including how each anticipated intercurrent event is handled. One of the available strategies, the treatment policy strategy, uses the value of the outcome variable regardless of whether the intercurrent event occurred. When applied to whether a patient continues treatment, the addendum notes that this "reflects the comparison described in the ICH E9 Glossary (under ITT Principle) as the effect of a treatment policy." That is the strategy closest to the classic reading of ITT.

The framework is considerably larger than this summary, with several attributes defining each estimand and several strategies for handling intercurrent events. Readers who want the full structure should go to the guideline itself; the European Medicines Agency publishes an accessible copy. The point to carry forward is narrower: the label describes an approach, while the estimand describes the question.

Discontinuing treatment is not the same as leaving the study

One distinction in the addendum resolves more confusion than any other, and it is routinely collapsed in casual reading.

ICH E9(R1) "distinguishes discontinuation of randomised treatment from study withdrawal. The former represents an intercurrent event, to be addressed in the precise specification of the trial objective through the estimand. The latter gives rise to missing data to be addressed in the statistical analysis."

These are different problems with different remedies. A participant who stops the drug but continues attending visits has generated an intercurrent event, and their outcome data exist. A participant who leaves the study entirely has generated missing data, and their outcome must be handled through the statistical analysis with the uncertainty that implies.

The order of operations matters here. As the addendum puts it, clarity in the estimand "gives a basis for planning which data need to be collected and hence which data, when not collected, present a missing data problem." Deciding what question you are asking comes first. It is a design decision, not a statistical repair carried out after the fact.

What CONSORT 2025 now asks authors to state

Reporting standards have caught up with this shift. The CONSORT 2025 statement, published in The BMJ on 14 April 2025, sets out a 30-item checklist for reporting randomized trials. Item 21, covering statistical methods, was completely revised, and item 21b now asks explicitly for the "Definition of who is included in each analysis (eg, all randomised participants), and in which group."

The practical consequence for a reader is direct. A paper is now expected to state the composition of each analysis rather than gesture at it. The bare sentence "analysis was by intention to treat" no longer satisfies the standard, because it does not tell you who was actually in the denominator.

That gives a short and honest checklist for reading a results section:

• How many participants were randomized, and how many were analyzed for the primary outcome?
• If those two numbers differ, does the paper explain the difference?
• Were participants who discontinued treatment still followed and assessed?
• How were missing outcomes handled, and does the paper test whether that choice changed the conclusion?

Where a trial report answers those four questions plainly, the reader can judge the analysis. Where it does not, the label on its own carries very little information.

How to read the phrase from now on

The most useful adjustment is to stop treating intention-to-treat as a quality badge and start treating it as a statement about which question was asked.

A trial reporting an ITT analysis is telling you it estimated the effect of assigning a treatment under a policy, with all the real-world imperfection that assignment entails. A trial reporting a per-protocol analysis is telling you something narrower and more fragile. Neither is inherently superior, and a well-conducted trial often reports both, alongside sensitivity analyses that test whether the conclusion survives different assumptions about the data that are missing.

What deserves scrutiny is not the label but the correspondence between the question stated and the data collected. That is where the reporting standards have moved, and it is a more demanding place to stand.

Readers interested in adjacent questions of study design may find the companion explainers on case-control versus cohort studies and relative versus absolute risk useful. A fuller list of methodological and epidemiological work is available on the peer-reviewed publications page.