Intent-to-Treat (ITT) vs. Per-Protocol Analysis

What counts as an intent-to-treat (ITT) analysis

ICH E9 (“Statistical Principles for Clinical Trials,” the harmonized guideline jointly adopted by FDA, EMA, and Japan’s PMDA) sets out the intention-to-treat principle: the primary analysis of a randomized trial should include all randomized subjects in the treatment group to which they were originally assigned, regardless of whether they met the entry criteria, regardless of the treatment they actually received, and regardless of any subsequent withdrawal from treatment or deviation from the protocol. In practice, because complete follow-up on literally every randomized subject is rarely achievable, ICH E9 Section 5.2 defines the Full Analysis Set (FAS) as the population that comes “as close as possible” to that ideal, permitting only minimal, pre-specified exclusions (for example, subjects with no post-randomization data at all). “ITT analysis” and “FAS analysis” are used near-interchangeably in trial reporting, though FAS is the more precise regulatory term.

The mechanism that makes ITT valuable is what it refuses to do: it never lets anything that happens after randomization — a participant stopping the drug, missing doses, switching arms, or dropping out — determine which group’s outcome data their results are counted in. Group membership is fixed at the moment of random allocation and stays fixed through analysis.

What counts as a per-protocol (PP) analysis

A per-protocol analysis instead restricts the analysis population to the Per-Protocol Set (PPS): the subset of randomized (and generally FAS-eligible) participants who received the assigned treatment substantially as specified, completed the relevant follow-up, and had no major protocol deviations — for example, no eligibility violations, no clinically important dosing deviations, and no use of a prohibited concomitant treatment. Where ITT asks “what happens if this treatment is assigned,” per-protocol asks “what happens if this treatment is actually taken as directed” — a narrower, more explanatory question about the biological effect of the treatment itself, isolated as far as possible from real-world non-adherence.

Why the distinction matters — and why “which is more conservative” isn’t a fixed rule

Which analysis gives the more cautious, harder-to-inflate result depends on the trial’s objective, and this is the part most often oversimplified:

• In a superiority trial (does the new treatment beat placebo or an active comparator?), ITT is normally the conservative, primary analysis. Non-adherence and dropout tend to dilute a real treatment effect toward the null in both arms rather than systematically favoring the new treatment, so ITT tends to underestimate efficacy relative to per-protocol. A per-protocol analysis, by discarding the participants who didn’t adhere, can inflate the apparent effect — which is exactly why ICH E9 recommends the Full Analysis Set as the primary analysis for superiority trials, with per-protocol run as a supportive sensitivity analysis, not the headline result.
• In a non-inferiority trial (does the new treatment perform no worse than an existing one, within a pre-specified margin?), this logic flips. Non-adherence and poor trial conduct tend to make two treatments look more similar to each other — diluting a true difference in both arms toward the null — which can bias an ITT analysis toward falsely concluding non-inferiority even when a real difference exists. A per-protocol analysis, by isolating participants who actually received the treatments as intended, is more likely to reveal a genuine difference if one exists, making it the analysis that’s harder to pass on a false-negative basis in this specific context. This is why regulatory guidance and the CONSORT extension for reporting non-inferiority and equivalence trials (Piaggio et al., JAMA, 2012) both recommend running ITT/FAS and per-protocol as co-primary analyses for a non-inferiority claim, requiring both to point the same direction before the claim is accepted — rather than treating either one alone as sufficient, or assuming ITT is automatically the safer choice the way it is in a superiority trial.

The safety literature adds a further wrinkle worth knowing rather than assuming away: at least one systematic review of antibiotic non-inferiority trials found ITT was more conservative than per-protocol in practice more often than the reverse — a reminder that “which population is conservative” is an empirical, trial-specific question, not something to state as a universal rule in either direction.

Relationship to randomization and the RCT

This whole distinction only exists because the underlying study is a randomized controlled trial. Randomization removes systematic differences between arms only at the single moment of allocation — it says nothing about what happens afterward. ITT is the analysis convention that protects randomization’s bias-control guarantee all the way through to the final analysis, by treating “assigned to group A” as the permanent, unchangeable definition of group membership. Per-protocol analysis allows post-randomization behavior (who adhered, who didn’t) to determine which participants’ data counts toward each arm’s result — which is exactly the kind of post-allocation selection that randomization was designed to prevent, and why per-protocol results are read as informative-but-not-definitive rather than as the trial’s primary evidence in most contexts.

Reporting quality for this distinction is itself a checklist item: CONSORT 2010 requires trials to state which analysis population(s) were used for each reported outcome and how participants were handled if excluded, precisely so that readers can tell whether “the trial worked” depended on which population was analyzed.

How this shows up in practice

A trial’s statistical analysis plan (SAP) pre-specifies both the Full Analysis Set and the Per-Protocol Set before unblinding, along with the criteria that place a given participant’s data in or out of each — see Clinical Trial Registration and Reporting Compliance for how these pre-specification requirements connect to trial registries and results reporting more broadly, and What Is a Clinical Trial? The NIH Definition Explained for the underlying study-design definition this population choice sits on top of.