Intention-to-Treat Analysis Explained

Intention-to-treat (ITT) analysis means every participant is analyzed in the group they were randomized to, regardless of whether they actually received, adhered to, or completed the assigned treatment. The common shorthand for what this “protects against” is usually stated too loosely — ITT is not primarily about being conservative, and it is not a guarantee that a trial’s headline result is trustworthy. What it specifically protects is the comparability that randomization created between groups. The moment you start moving participants out of their randomized group — because they stopped the drug, crossed over, or didn’t meet an eligibility criterion discovered after enrollment — you are re-introducing the exact selection process randomization was designed to eliminate, and the two groups stop being exchangeable. Per-protocol (PP) analysis does exactly that: it restricts the comparison to participants who adhered to the protocol as specified, which is precisely the subset most likely to differ systematically between arms.

What intention-to-treat actually protects against

Randomization balances measured and unmeasured characteristics between groups at the moment of assignment. It does not, and cannot, guarantee balance after assignment if people are then excluded on the basis of something that happened after randomization — non-adherence, side effects, disease progression, or a clinician’s judgment that a participant no longer fits the protocol. All of those post-randomization events can be related to prognosis. A patient who stops a chemotherapy regimen because the disease is progressing too fast to tolerate it, and a patient who stops because they experienced only a mild, unrelated rash, are not interchangeable — but a per-protocol analysis that simply drops “non-completers” can end up dropping both, and dropping different proportions of each type from each arm.

ITT analysis protects against this by keeping the comparison anchored to randomization: everyone randomized is counted in the arm they were assigned to, and — critically — the reason they left the protocol is not used to decide whether they’re in the analysis. This is why ITT is described as testing the effect of being assigned to a treatment strategy, not the effect of actually receiving it. In ICH E9 (“Statistical Principles for Clinical Trials”), this is formalized as the intention-to-treat principle, and because a literal zero-exclusion, complete-follow-up ITT population is rarely achievable in practice, ICH E9 defines the Full Analysis Set (FAS) — the population as close as possible to all-randomized, with only minimal, pre-specified exclusions (for example, participants later found to have no data at all post-randomization). In most trial reporting, “ITT” and “FAS” are used near-interchangeably; FAS is the more precise regulatory term.

Worked example: the same trial, read two ways

Illustrative example — this is a constructed teaching scenario, not a report of a real trial or its results. Picture a two-arm superiority trial of a new oral therapy against standard care, 200 participants randomized to each arm, primary outcome a continuous biomarker measured at 24 weeks. In the treatment arm, 40 participants stop the drug before week 24 because of a tolerability problem that is, unsurprisingly, correlated with poorer response — participants who were going to benefit less were also more likely to find the side effects not worth tolerating. In the control arm, 15 participants are lost to the analysis for reasons unrelated to prognosis (relocation, withdrawal of consent for unrelated reasons).

Analysis set	Treatment arm (n analyzed)	Control arm (n analyzed)	Mean improvement, treatment	Mean improvement, control	Between-group difference (95% CI)
Intention-to-treat / FAS	200	200	8.0 units	4.5 units	3.5 (1.6 to 5.4)
Per-protocol	160	185	10.5 units	4.7 units	5.8 (3.7 to 7.9)

The per-protocol estimate is larger — not because per-protocol analysis is more sensitive, but because it removed the treatment-arm participants who were disproportionately the poor responders, while removing a roughly random slice of the control arm. The sentence a researcher should write about the ITT/FAS result: “In the full analysis set, treatment was associated with a 3.5-unit greater improvement than control (95% CI 1.6 to 5.4), reflecting the effect of the assigned treatment strategy including observed non-adherence.” The per-protocol figure of 5.8 is not a “more accurate” estimate of drug efficacy — it is an estimate contaminated by exactly the kind of selection that non-adherence correlated with prognosis produces, and it should be reported and interpreted as a supportive, hypothesis-generating comparison, not the trial’s primary claim.

Why per-protocol analysis reintroduces selection bias

The mechanism is the same one that produces confounding and selection bias anywhere else in observational research — the only difference is that here it is created by the analysis choice rather than the original study design. Once you condition the comparison on something that happened after randomization (completed the protocol vs. did not), and that “something” is itself related to the outcome, the two groups being compared are no longer the two groups randomization created. This is structurally identical to conditioning on a post-treatment variable in any causal analysis: it can open a biasing path between treatment and outcome even though none existed at baseline. It’s also why “as-treated” analysis — grouping participants by the treatment they actually received rather than the one they were assigned — has the same vulnerability in a sharper form, since it discards randomization entirely as the basis for group membership.

This does not make per-protocol analysis illegitimate. It makes it a different question with a different, more fragile evidentiary basis: not “what happens if this treatment strategy is adopted,” but “what happens among people who can and do adhere to it as specified” — a question that is useful, but that requires much stronger assumptions (essentially, that adherence itself is unrelated to unmeasured prognosis, an assumption randomization does nothing to support once you’re restricting on post-randomization adherence).

When per-protocol analysis is the right supporting analysis

As a sensitivity analysis alongside ITT/FAS. ICH E9 recommends the Full Analysis Set as the primary analysis for superiority trials and per-protocol as a supportive analysis — if both point the same direction and similar magnitude, that agreement strengthens confidence in the result; if they diverge sharply, the divergence itself is informative about how much adherence is driving the effect.
For biological/mechanistic questions where the object of interest is genuinely “what does the drug do when taken as directed” rather than “what happens when this strategy is deployed in practice” — for example, early pharmacodynamic or dose-finding work, though even there the limitation above still applies.
For safety analyses, which are conventionally handled differently again — typically analyzed by treatment actually received (the “safety population”), since attributing an adverse event to a drug a participant never took would be a different kind of error.

The non-inferiority inversion

Everything above describes superiority trials, where the concern is a treatment falsely appearing better than it is. In a non-inferiority trial, the usual bias direction flips, and this is the single most common place ITT-vs-per-protocol guidance gets misapplied. Non-adherence and poor trial conduct tend to dilute a real difference between treatments toward the null — pulling both arms’ results closer together. In a superiority trial, that dilution works against the sponsor’s hypothesis, so it’s a conservative, safe direction to err in. In a non-inferiority trial, the hypothesis being tested is that the treatments are similar, so a bias that pulls results toward “no difference” makes it easier to falsely conclude non-inferiority even when a real, clinically important difference exists. ITT/FAS analysis, on its own, is no longer the automatically conservative choice.

Because of this, the CONSORT extension for reporting non-inferiority and equivalence trials (Piaggio et al., JAMA 2012) and regulatory guidance both call for ITT/FAS and per-protocol to be run as co-primary analyses for a non-inferiority claim, with the conclusion considered robust only if both analyses agree. Per-protocol, by isolating participants who actually received treatment as intended, is generally more likely to reveal a genuine difference in this specific setting — though this is a tendency, not an absolute rule: at least one systematic review of antibiotic non-inferiority trials found ITT was empirically more conservative than per-protocol more often than the reverse, which is a real, citable caution against treating “which analysis is conservative” as fixed in either direction. The safe practice is procedural, not directional: pre-specify both analyses, pre-specify the non-inferiority margin and its justification, and require agreement rather than picking whichever result supports the claim after the fact.

Assumptions and limits — what ITT does not fix

Treating “we ran ITT” as a completed checkbox misses several things ITT does not, by itself, solve:

It does not solve missing outcome data. ITT specifies who is analyzed and in which group — it says nothing about what to do when a participant’s outcome was never observed. That’s a separate, and often larger, source of bias, handled through pre-specified missing-data strategies (multiple imputation, mixed models for repeated measures, or, historically and less defensibly, last-observation-carried-forward) and reported sensitivity analyses under different missingness assumptions.
It is only as good as the randomization it protects. If the underlying randomization was compromised — inadequate allocation concealment, or a placebo-run-in period that excluded poor responders before random assignment — ITT preserves the comparability of a flawed starting point, it doesn’t repair it.
“Modified ITT” is a heterogeneous, not a standardized, term. Trials frequently report a “modified intention-to-treat” (mITT) population — most often all-randomized-who-received-at-least-one-dose. This is a legitimate, common design choice, but because the exact exclusion rule varies by trial, it must be defined explicitly in methods, not left for a reader to assume matches another trial’s mITT definition.
Direction of bias is context-dependent, not a fixed law. The “ITT is conservative” heuristic holds reasonably well for superiority trials with the classic dropout-correlates-with-poor-response pattern; it does not hold universally, as the non-inferiority case above shows, and it can also fail in superiority trials where non-adherence happens to correlate with better rather than worse prognosis.

How to write the sentence in a results section

The reporting language should make the analysis set and its implication explicit rather than just naming it:

“The primary analysis was intention-to-treat, including all randomized participants in their assigned group regardless of adherence; between-group difference was 3.5 units (95% CI 1.6–5.4). A pre-specified per-protocol sensitivity analysis, restricted to participants completing treatment as assigned, showed a larger difference (5.8 units, 95% CI 3.7–7.9), consistent with treatment-arm attrition being concentrated among lower responders; the ITT estimate is reported as the primary result.” That single sentence does three things a bare “ITT was used” does not: it states which population was primary, it reports the alternative rather than hiding a divergence, and it gives the reader the mechanism rather than leaving them to assume the two numbers should have matched.

Frequently asked questions

What does intention-to-treat analysis actually mean?

Every participant is analyzed as a member of the group they were randomly assigned to, regardless of whether they received, adhered to, or completed that assignment — including participants who stopped treatment, crossed over, or turned out not to meet an eligibility criterion. See the Intent-to-Treat vs. Per-Protocol Analysis dictionary entry for the formal definitions side by side.

Is intention-to-treat always the more conservative analysis?

No. It tends to be conservative in superiority trials, where non-adherence dilutes a real effect toward the null in a direction that works against the hypothesis being tested. In non-inferiority trials the same dilution works in the opposite direction — toward falsely supporting the “no meaningful difference” hypothesis — which is why regulatory guidance requires ITT and per-protocol to be run as co-primary analyses there, not ITT alone.

Why does excluding non-adherent participants bias the comparison?

Because non-adherence is frequently related to prognosis (participants who tolerate a drug poorly, or whose disease is progressing fastest, are more likely to leave the protocol), excluding them is equivalent to conditioning the comparison on a post-randomization variable connected to the outcome. That reintroduces a selection process between groups that randomization was specifically designed to remove.

What is the Full Analysis Set, and how is it different from strict ITT?

The Full Analysis Set (FAS), defined in ICH E9, is the population “as close as possible” to a literal all-randomized ITT population, with only minimal, pre-specified exclusions (for example, participants with no post-randomization data at all). A completely exclusion-free ITT population is rarely achievable in practice, so FAS is the term regulatory statisticians use for what most published papers mean when they say “ITT.”

Is per-protocol analysis ever the better choice for a primary result?

Rarely as the sole primary analysis in a randomized superiority trial, because it forfeits the protection randomization provides. It is standard and appropriate as a pre-specified supporting or sensitivity analysis, and it becomes co-primary — not secondary — for non-inferiority and equivalence trials, per the CONSORT non-inferiority/equivalence extension.

What is “modified intention-to-treat” (mITT)?

A commonly used but non-standardized variant, most often defined as all randomized participants who received at least one dose of the assigned treatment. Because the precise exclusion rule differs across trials, an mITT population is only interpretable if the paper defines exactly which participants it excludes and why.

Intention-to-Treat Analysis: What It Protects Against and When Per-Protocol Misleads

Ask about Intention-to-Treat Analysis: What It Protects Against and When Per-Protocol Misleads