Intention-to-Treat, Full Analysis, and Per-Protocol Sets

“Analyzed by intention to treat” is not enough information to understand a trial result. Readers still need to know how many participants were randomized, received an assigned intervention, remained in follow-up, had the outcome observed, and entered the reported analysis. They also need the clinical question, treatment effect being estimated, exclusions, missing-data assumptions, and sensitivity analyses.

Intention-to-treat is a principle tied to randomized assignment. A full analysis set is an operational attempt to approach that ideal. A per-protocol set is a defined subset. None of those labels is an imputation method, and none identifies the target treatment effect as precisely as an estimand.

The intention-to-treat principle

ICH E9 states that the intention-to-treat principle implies a primary analysis including all randomized participants.[1] Under the ideal, participants are followed, assessed, and analyzed according to the group to which they were randomized, regardless of later compliance.

Preserving initial randomization matters because assignment by chance provides a secure foundation for comparing groups. Excluding people after randomization can make the groups differ for reasons related to prognosis, response, adverse events, or adherence. The resulting comparison may no longer have the protection created at assignment.

The ideal also requires outcome follow-up after intervention discontinuation when those data are relevant to the question. Assignment does not disappear when a participant stops the assigned intervention. Study withdrawal and intervention discontinuation are different events and should be recorded separately.

A full analysis set approximates the ideal

ICH E9 uses full analysis set for the set that is as complete as possible and as close as possible to the ITT ideal. It identifies limited circumstances that may lead to exclusions, including major eligibility violations, no exposure to the assigned intervention, or no post-randomization data. These are not automatic permissions to remove participants; every exclusion and its potential bias need justification.[1]

For eligibility exclusions, E9 describes four safeguards: the criterion was measured before randomization; violations can be detected objectively; every participant receives equal scrutiny; and all detected violations of that particular criterion are excluded. Blinded review matters because knowledge of assignment or outcomes can affect which problems are noticed.[1]

Similarly, excluding someone who never started the intervention requires attention to why that happened. E9 gives the example that the decision to begin could not have been influenced by knowledge of the assigned intervention. “Never treated” is not, by itself, proof of a harmless exclusion. Lack of post-randomization data also requires examination of potential bias; it is not a way to make incomplete follow-up disappear.[1]

Calling an analysis “full analysis” does not establish that it contains everyone randomized. The report should state the exact rule and group counts. ICH E9 does not define “modified intention to treat.” In published reports, that modifier may accompany exclusions based on receipt of an intervention, availability of post-baseline measurements, eligibility, or another rule. The report must therefore state the exact operational definition rather than rely on the label.

An analysis set is also not a missing-data solution. A participant can belong to the full analysis set yet lack the outcome at a particular time point. A model, weighting method, imputation rule, or other assumption may be used to estimate the result. That choice needs its own description and sensitivity analysis. See the estimands and missing-data guide for the distinction.

What a per-protocol set selects

ICH E9 describes the per-protocol set as a subset of the full analysis set characterized by greater compliance with the protocol.[1] Criteria may involve sufficient exposure under the protocol, availability of the primary outcome, and absence of defined major protocol violations. E9 says that precise reasons for exclusion should be fully defined and documented before breaking the blind, in a manner appropriate to the trial. Analysis-set rules belong in the protocol, and any permitted refinements during blinded review need a record; a favorable result is not a reason to redefine the set afterward. The registration, protocols and analysis-plans guide explains why version and timing matter.

Per-protocol does not mean “participants for whom the intervention truly works.” It selects people based on events occurring after randomization. Adherence, discontinuation, missing outcomes, or a major deviation may be related to intervention assignment and prognosis. Removing participants on that basis can introduce selection bias, even when the same number is removed from each group.

Nor is a conventional per-protocol comparison automatically an unbiased estimate of the effect of adherence. Estimating adherence effects can require methods addressing post-randomization prognostic factors that influence both continued adherence and outcomes. A simple subset analysis may not provide that adjustment.

E9(R1) goes further: a conventional per-protocol analysis may not align with a relevant estimand at all. It does not automatically estimate the effect in a principal stratum, such as people who would tolerate an intervention. Principal stratification concerns potential intercurrent events under the alternative assignments, not simply the participants observed to adhere in each assigned group. Those observed subsets may contain different kinds of people.[2]

Analysis sets and estimands answer different questions

ICH E9(R1) distinguishes the target, method, result and roles of additional analyses:[2]

An estimand specifies the treatment conditions, population, variable or endpoint, population-level summary, and how relevant events after treatment begins affect the question. Intercurrent events affect interpretation or existence of the measurements of interest. They may include intervention discontinuation, use of an additional intervention, switching, or a terminal event when that event is not already part of the outcome definition.[2]

Protocol deviations and intercurrent events are not synonyms. A visit outside a required window can be a deviation without necessarily being an intercurrent event. Conversely, death or permitted rescue treatment can affect the question without representing a protocol violation. Study withdrawal or administrative censoring is not, in itself, an intercurrent event; it can create missing data for the specified estimand.[2]

The analysis population contributes to implementation, but naming a set does not fully specify those attributes. Two analyses can use the same people yet target different questions. Conversely, different estimators may target the same estimand under different assumptions.

Treatment-policy, while-on-treatment, and hypothetical questions

ICH E9(R1) describes several strategies for intercurrent events.[2] Three commonly encountered strategies illustrate why “ITT versus per protocol” is too narrow:

• Under a treatment-policy strategy, outcomes are used regardless of whether the specified intercurrent event occurs. When applied to discontinuation or changes in other interventions, this can reflect a question about assignment under the policies being compared. Relevant follow-up after that event is therefore important; the label does not supply unobserved outcomes.
• Under a while-on-treatment strategy, the observation period of interest ends at the intercurrent event. Interpretation requires care if the timing or frequency of that event differs between groups.
• Under a hypothetical strategy, the question concerns outcomes in a clearly defined scenario in which the intercurrent event would not occur. Estimation depends on assumptions because some required outcomes were not observed in that scenario. The scenario needs clinical justification; simply assuming everyone could have continued despite toxicity is not automatically an informative question.

These are different questions, not competing ways to manipulate the same answer. A trial can use different strategies for different intercurrent events. The choice should follow the objective, be specified in advance, and align with data collection and analysis.

Treatment policy generally cannot handle a terminal event by using a measurement that no longer exists, for example, a functional score after death. The endpoint and strategy must address that situation explicitly. A composite strategy may incorporate the event into the outcome, while a while-alive strategy changes the period of interest; neither is equivalent to inventing a post-death measurement. E9(R1) also describes principal-stratum strategies, which are distinct from an observed per-protocol subset.[2]

Missing data remain separate. Data intended to inform the estimand but not collected are missing. Measurements that do not exist or are not meaningful because of the chosen intercurrent-event strategy are conceptually different.[2]

Sensitivity is not just a second analysis

A sensitivity analysis asks whether inference about the same estimand changes when important assumptions or data limitations are varied. Switching both the population and the clinical question, then obtaining a similar P value, is not automatically evidence that the original inference is robust.[2]

Supplementary analyses can provide useful additional insight, but they have a different role. E9(R1) revisits the older advice to compare full-analysis and per-protocol results: agreement may be reassuring, yet a biased subset or an analysis without a relevant target does not become a valid sensitivity test merely because it agrees. Differences need explanation in terms of participants, observations, target effect and assumptions.[1,2]

An analysis plan should identify the main estimator and suitable sensitivity analyses in advance. Reports should distinguish planned analyses, changes made during blinded review and post-hoc exploration. No imputation rule is universally valid, and repeatedly running plausible-looking models does not establish that untestable assumptions are true.

The noninferiority and equivalence caveat

In many superiority trials, including noncompliant participants according to assigned groups can attenuate an effect, making a favorable difference harder to demonstrate. This observation is sometimes turned into the false rule that ITT is always conservative.

ICH E9 explicitly warns that, in equivalence or noninferiority trials, full-analysis-set use is generally not conservative and its role requires careful consideration.[1] If discontinuation, switching, additional intervention, nonadherence, or poor conduct makes group outcomes look more similar, attenuation can support an incorrect conclusion of noninferiority or equivalence.

ICH E9(R1) preserves that warning for treatment-policy estimands in these settings.[2] FDA’s final 2016 Non-Inferiority Clinical Trials guidance addresses interpretable design, conduct, margins, and analysis for regulatory submissions.[4] It is non-binding guidance, not a general permission to choose whichever population gives the desired conclusion.

Full-analysis and per-protocol results may both be reported in noninferiority work, but neither label settles validity. Agreement does not cure a weak margin, absent assay sensitivity, extensive deviations, shared missing-data assumptions or a mismatch in the target effect. The Noninferiority, Equivalence, and Superiority guide covers those design-specific boundaries.

Report participant flow before labels

For each assigned group, a useful study summary states:

1. how many were randomized;
2. how many received any assigned intervention;
3. how many discontinued that intervention and why;
4. how many remained in study follow-up;
5. how many had the outcome observed at the stated time;
6. how many entered each named analysis set; and
7. how many contributed to the estimate under the stated missing-data method.

Outcome-specific denominators matter. A participant may contribute to one time point or endpoint but not another. Percentages without the correct denominator can hide attrition. Differences between randomized, treated, followed, observed, and analyzed counts are not editorial clutter; they define what the result represents.

Equal numbers excluded or lost from each group do not establish absence of bias. Reasons, timing, relationship to the outcome and the assumptions used for unobserved data still matter. Read these counts alongside the effect estimate and its confidence interval, not as a substitute for uncertainty or source-level appraisal.

Important limitations

• The ITT principle is an ideal tied to randomized assignment; the label alone does not prove complete follow-up or inclusion.
• Full analysis sets may have limited justified exclusions, but no-dose or missing-data labels alone do not establish that exclusions were harmless.
• ICH E9 does not define “modified ITT”; every report using the term needs an exact operational definition.
• Analysis-set membership and missing-data handling are separate decisions.
• Conventional per-protocol subsets can be biased, may not align with a relevant estimand, and are not automatically principal-stratum analyses.
• Sensitivity analyses target the same estimand; other supplementary analyses may answer different questions.
• Treatment-policy language does not create a meaningful endpoint value after a terminal event.
• ITT-like and full-analysis approaches are not automatically conservative in noninferiority or equivalence trials.
• This page explains trial interpretation and does not recommend treatment or use of any intervention.

Frequently asked questions

Does ITT mean every randomized participant had a measured outcome?

No. ITT describes an analysis principle. A report must separately disclose follow-up, observed outcomes, missing data, and how any unobserved values affected estimation.

Is a full analysis set identical to all randomized participants?

Not necessarily. It is intended to be as complete and as close to the ITT ideal as possible, but limited exclusions can occur and require justification.

What does modified ITT mean?

ICH E9 does not define the term. Read the study’s exact rule and report who was excluded, when, why, and whether the rule was prespecified.

Is imputation part of the ITT principle?

No. Imputation or modeling is an assumption-dependent missing-data method. It is distinct from the principle and from membership in an analysis set.

Is per protocol automatically less biased?

No. Selection after randomization can depend on assignment, prognosis, response, or adverse events. A naïve subset comparison can introduce substantial bias.

Is ITT always conservative?

No. Especially in noninferiority and equivalence trials, attenuation of group differences can make an ITT-like or full-analysis result anti-conservative for the study objective.

Which denominator should a study summary use?

It should state all relevant denominators: randomized, treated, followed, observed for the outcome and time point, included in each analysis set, and contributing to the estimate.

Primary and official sources

1. ICH E9: Statistical Principles for Clinical Trials
2. ICH E9(R1): Estimands and Sensitivity Analysis in Clinical Trials
3. ICH E9(R1): ICH-Hosted Training Material (Not Official Guidance or Policy)
4. FDA: Non-Inferiority Clinical Trials, Final Guidance, including the November 2016 guidance PDF

Review and corrections

Editorial evidence review: September 3, 2026, in both America/Edmonton and UTC. Serum Research is responsible for this source-based educational guide. The review checked the official ICH guidance and relevant FDA material; it is not an independent clinical or statistical specialist review, a new analysis of a named trial, or an assessment of any intervention's suitability for use.

The sources have different roles: ICH E9 is dated February 5, 1998; the E9(R1) final addendum was adopted November 20, 2019; the December 2021 training material expressly represents its authors' views, not additional official guidance or policy. This review used its summary module, not all 238 pages. The FDA source is final, nonbinding November 2016 guidance.[1–4]

See how we review evidence, return to the Research Library or study records, or report a source discrepancy through Contact Us.