Scientific Writing

Intention-to-Treat Analysis vs Per-Protocol: What to Report

DE By Directive Editorial Team, Directive Publications ·16 Nov 2026 ·7 min read
Intention-to-Treat Analysis vs Per-Protocol: What to Report

Report an intention-to-treat analysis as the primary result of a superiority trial, because it keeps the comparison that randomisation created. Present a per-protocol analysis beside it, defined before unblinding, as a check on how adherence changes the answer. In non-inferiority trials, give both equal weight: a conclusion is more convincing when both analyses support it.

The definition, and the rule that excluding any randomised participant ends strict intention-to-treat (ITT), are set out in labelling the analysis box of a CONSORT flow diagram. CONSORT 2025 item 21b asks who each analysis includes; the wording is in what changed in CONSORT 2025. This guide covers why the two analyses answer different questions, and how to write both up.

Why intention-to-treat analysis protects the randomised comparison

Randomisation makes the groups comparable at the moment of allocation. Everything that happens afterwards, such as stopping treatment, missing sessions or switching groups, is a response to the treatment and to the patient's own prognosis. An intention-to-treat analysis compares the groups as randomised, so it estimates the effect of assigning a treatment in the way the trial delivered it.

A naive per-protocol (PP) analysis keeps only participants who adhered. People who tolerate a treatment, attend every visit or feel well enough to continue can differ in prognosis from those who stop. The PP comparison is therefore between groups chosen by behaviour after randomisation, and it can be confounded in the same way as an observational comparison.

A worked example: the two estimates diverge

The trial below is invented to illustrate the structure; its numbers are not real data. Two hundred adults per group were randomised to a supervised rehabilitation programme or usual care, and recovery at 12 weeks was recorded for everyone. In the programme group, 60 people stopped attending; they had more severe disease at baseline. In the usual care group, 10 had a major protocol deviation.

AnalysisWho is includedProgrammeUsual careDifference, percentage points (95% CI)
Intention-to-treatAll 400 randomised, as allocated100/200 (50%)80/200 (40%)10.0 (0.3 to 19.7)
Per-protocolAdherent, no major deviation84/140 (60%)76/190 (40%)20.0 (9.3 to 30.7)
Excluded from per-protocolStopped or deviated16/60 (27%)4/10 (40%)Not a comparison

The PP estimate doubles. The programme did not work twice as well in adherers; the analysis removed the people with the worst prognosis from one arm only. The 95% confidence intervals (CIs) are normal approximations.

Name the estimand before choosing intention-to-treat or per-protocol

An estimand is a precise statement of the treatment effect a trial sets out to estimate. The International Council for Harmonisation (ICH) E9(R1) addendum on estimands, adopted in November 2019, builds it from five parts: the treatment, the population, the outcome variable, how intercurrent events are handled, and the population-level summary.

Intercurrent events are events after treatment starts that affect how outcome measurements are interpreted, or whether they exist. The addendum's examples include discontinuation of assigned treatment, use of an additional or alternative therapy (such as rescue medication or switching) and death. For each one, you choose a strategy.

ICH E9(R1) strategyQuestion it answersRelation to the familiar labels
Treatment policyWhat is the effect of assignment, whatever happens after the event?Closest to the intention-to-treat principle; needs outcomes collected after discontinuation
HypotheticalWhat would the effect be if the event, such as rescue medication, had not occurred?One way to frame an adherence question with a stated scenario
Composite variableIs the event itself part of the outcome, for example counted as treatment failure?Keeps everyone randomised in the analysis
While on treatmentWhat is the response before the event occurs?Must say what happened after the event as well
Principal stratumWhat is the effect within a principal stratum, for example people who could tolerate the test treatment?Not what a naive per-protocol analysis estimates

The addendum is blunt about the last row. It says analysis of the per protocol set does not estimate the effect in any principal stratum, because it may not compare similar subjects, and that a relevant estimand to match it may not exist. It also says a PP set result "might be subject to severe bias".

Non-inferiority trials: why intention-to-treat analysis is not automatically conservative

In a superiority trial, non-adherence and switching tend to make the groups look alike, which works against showing a difference. In a non-inferiority or equivalence trial, the same dilution works in favour of the claim being tested. ICH E9 therefore describes the full analysis set result as generally not conservative in these trials, and the addendum extends the point to treatment-policy estimands.

For these designs, plan and report both analyses against the same prespecified margin. ICH E9 says it is usually appropriate to plan analyses on both the full analysis set and the per protocol set, and that consistent results increase confidence. Report the trial with the CONSORT extension for non-inferiority and equivalence trials, an extension built on the earlier CONSORT checklist and not yet realigned with CONSORT 2025. When the two analyses disagree, do not choose the favourable one. Say that non-inferiority was shown in one analysis and not the other, give both CIs against the margin, and discuss which intercurrent events drove the difference.

Define the per-protocol population before unblinding

A per-protocol population is only credible if its rules were fixed before anyone saw results by group. Write three things into the statistical analysis plan: a minimum adherence threshold, a closed list of major protocol deviations, and how switchers are handled. ICH E9(R1) notes that a protocol deviation, such as a visit outside its time window, is not the same as an intercurrent event; list each separately.

An as-treated analysis is not a per-protocol analysis. As-treated moves switchers into the group whose treatment they received. Per-protocol excludes them. Both break randomisation, and neither should be labelled primary in a superiority trial. Missing outcomes are a separate problem with their own methods, covered in choosing and justifying a missing data method.

Copy this paragraph into your Methods, replace every bracketed slot and delete what does not apply.

Analysis populations. The primary analysis followed the intention-to-treat principle: all [number] randomised participants were analysed in the group to which they were allocated. Missing outcomes were handled by [method]. The per-protocol population was defined in the statistical analysis plan ([version], [date]) before unblinding. It excluded participants who received less than [minimum adherence, for example 8 of 12 sessions] or had a major protocol deviation, listed in advance as [list]. Participants who switched to the other group's intervention were analysed as randomised in the primary analysis and excluded from the per-protocol population. Estimand. The primary estimand addressed [intercurrent event] with a [treatment policy] strategy and [intercurrent event] with a [hypothetical / composite variable / while on treatment] strategy. Supplementary analysis. Adherence-adjusted effects were estimated by [complier average causal effect / inverse probability weighting], assuming [key assumption].

Report both analyses with denominators in the Results and abstract

Put the primary and per-protocol results in one table, with the number analysed in each group, the effect size and its CI, as in the worked example. State which analysis is primary in the table title, the Results text and the abstract. If the abstract has room for only one result, it is the primary one, followed by a phrase saying whether the per-protocol result agreed.

If the question is the effect of actually receiving treatment, a naive PP analysis is the weakest tool. A complier average causal effect (CACE) analysis uses randomisation as an instrument to estimate the effect in people who would adhere if offered treatment. Inverse probability weighting (IPW) reweights adherers to resemble the full randomised group, using recorded predictors of adherence. Both rest on assumptions that need stating, so present them as supplementary or sensitivity analyses. Effects in groups defined after randomisation raise the same problem discussed in reporting a subgroup analysis without overclaiming.

What reviewers query in intention-to-treat and per-protocol reporting

  • A per-protocol population whose rules appear only in the paper, not in a dated plan written before unblinding.
  • An as-treated analysis presented as the primary result of a superiority trial.
  • Intention-to-treat and per-protocol results that disagree, with only the favourable one in the abstract.
  • A non-inferiority claim resting on one analysis set, with no comment on the other.
  • An adherence threshold chosen after looking at outcomes by adherence level.
  • No estimand stated, so a reader cannot tell what question the primary result answers.

Systematic reviewers meet the same choice from the other side: the RoB 2 tool asks them to judge either the effect of assignment or the effect of adherence, as described in choosing a risk of bias tool.

At Directive Publications, the author guidelines ask you to match a trial manuscript to the relevant EQUATOR checklist, and our reviewer guidelines ask reviewers whether the statistics or analytical methods are appropriate and correctly applied. We have not published a separate statistical-review policy. If you find an error in this guide, report the problem to us.

Frequently Asked Questions

Can a per-protocol analysis be the primary analysis of a randomised trial?
In a superiority trial it should not be, because excluding people for behaviour after randomisation can confound the comparison. In a non-inferiority or equivalence trial, ICH E9 treats both analysis sets as important, and the conclusion is stronger when both support it. Whatever the design, name the primary analysis in the statistical analysis plan before unblinding.
Is an as-treated analysis the same as a per-protocol analysis?
No. An as-treated analysis moves participants who switched into the group whose treatment they actually received, while a per-protocol analysis excludes them and anyone below the adherence threshold. Both depart from the randomised comparison, so label each one by what it does and report it as a secondary analysis.
What is an intercurrent event in a clinical trial?
An intercurrent event is something that happens after treatment starts and affects how the outcome measurements are interpreted, or whether they exist. The ICH E9(R1) addendum gives examples such as stopping the assigned treatment, using rescue medication, switching treatment and death. Each one needs a stated strategy, such as treatment policy or hypothetical, when the estimand is defined.
What should a paper say when the intention-to-treat and per-protocol results disagree?
Report both results with their denominators and confidence intervals, and do not present only the favourable one in the abstract. Explain which intercurrent events or exclusions differ between the two populations and how they could move the estimate. In a non-inferiority trial, say plainly that non-inferiority was shown in one analysis but not the other.
DE
Directive Editorial Team
Directive Publications

The editorial team at Directive Publications — an international open-access publisher of peer-reviewed medical and scientific journals.

Publishing your research?

Directive Publications is an open-access publisher — every article peer-reviewed, Crossref-registered, and free to read under CC BY 4.0.

Submit a manuscript →Read our Scientific Writing policy →