Clinical Study Connect

Plain-Language Clinical Study and Ingredient Research

Per-Protocol Versus Intention-to-Treat Results: Two Views of One Trial

posted on

By ClinicalStudyConnect.com Research Desk

The Short Answer: Two Analyses, Two Different Questions

An intention-to-treat (ITT) result compares everyone who was randomized, counted in the group they were assigned to, whether or not they followed the plan. A per-protocol (PP) result compares only the participants who received their assigned treatment without major protocol deviations. The SPIRIT 2025 guidance for writing trial protocols explains that keeping all randomized participants in the analysis, each in their originally allocated group, is what preserves the benefits of randomization, and it describes the ITT population as widely recommended as the preferred analysis population (SPIRIT 2025 explanation, Item 27b).

So neither number is simply “the real result.” The ITT result tells you what happened to people assigned to a treatment under trial conditions. The per-protocol result tries to tell you what happened to people who followed it, but it gives up some of the protection that randomization provides. When the two agree, that is reassuring. When they disagree, the gap is information to interpret, not a reason to pick the more favorable number.

Why One Trial Can Produce Two Results

Randomization creates groups that are expected to be similar at the start. After randomization, things happen: some participants stop taking the intervention, some miss doses, some switch treatments, some are found to be ineligible, and some leave before their outcome is measured.

Each analysis handles those events differently:

  • ITT keeps everyone in their assigned group. The comparison stays a comparison of two randomized groups, but the effect estimate includes participants who took little or none of the intervention.
  • Per-protocol removes participants who deviated. The remaining groups are no longer the full randomized groups, and they may no longer be comparable.

The SPIRIT 2025 explanation states this risk directly: excluding data from protocol non-adherers can compromise randomization and introduce bias, particularly when how often and why people stop adhering differs between groups. It gives “healthy adherers” bias as an example (SPIRIT 2025, Item 27b).

Analysis Comparison Card: ITT Versus Per-Protocol

Intention-to-treat (ITT)

  • Who is included: All randomized participants. An example protocol quoted in the SPIRIT 2025 explanation spells this out as including those later found ineligible, those who did not adhere, and those never treated.
  • Which group they count in: The group they were randomized to, regardless of what they actually received.
  • Question it answers: What is the effect of being assigned to this intervention compared with the comparator, under the conditions of this trial?
  • Main strength: Preserves the comparability created by randomization.
  • Main limitation: If many participants did not take the intervention as planned, the result may not reflect what happens when it is taken as intended. Missing outcome data still has to be handled, and SPIRIT 2025 notes that filling in missing outcomes does not guarantee avoidance of bias except under strong assumptions.
  • Typical role: Recommended by SPIRIT 2025 as the preferred analysis population. The example protocol it quotes uses ITT for the primary outcome.

Per-protocol (PP)

  • Who is included: Randomized participants who received their allocated treatment without major protocol deviations. Each trial’s protocol or statistical analysis plan defines what counts as “major.”
  • Which group they count in: Their assigned group, but only if they stayed on protocol.
  • Question it attempts to answer: What is the effect among participants who followed the protocol?
  • Main strength: Speaks to a different question than ITT: what happens when the intervention is used as directed.
  • Main limitation: Adherent and non-adherent participants can differ in ways that affect outcomes, so a simple per-protocol comparison can reflect who adhered rather than what the intervention did.
  • Typical role: In the example protocol quoted by SPIRIT 2025, the per-protocol population is used as a sensitivity analysis alongside the primary ITT analysis.

Reading guidance when both are reported

  • Similar results: This suggests deviations did not change the conclusion. Both analyses can still share other weaknesses, such as a small sample, short duration, or outcome choice.
  • Per-protocol noticeably more favorable than ITT: Check adherence and dropout by group before accepting the larger effect. The gap can reflect real benefit among adherers, selection bias, or both.
  • Only a per-protocol or “completers” result is reported: Treat the finding with extra caution and look for the full randomized counts in the trial registry or flow diagram.

What Adherence Bias Looks Like in Real Trial Data

A widely cited illustration comes from the Coronary Drug Project, which evaluated several lipid-influencing drugs for long-term treatment of coronary heart disease. In its clofibrate comparison, five-year mortality was 20.0 percent in 1,103 men given clofibrate and 20.9 percent in 2,789 men given placebo (P = 0.55) (Coronary Drug Project Research Group, NEJM 1980).

Men who took at least 80 percent of their assigned clofibrate had much lower mortality than poor adherers: 15.0 versus 24.6 percent. On its own, that could look like evidence the drug worked when taken. But the same pattern appeared in the placebo group: 15.1 percent mortality among good adherers versus 28.3 percent among poor adherers. The investigators concluded that these findings show the serious difficulty of evaluating treatment efficacy in subgroups defined by how patients responded to the protocol after randomization, such as adherence (NEJM 1980 abstract).

The takeaway for readers: people who adhere to a study protocol can differ from those who do not in ways that affect outcomes. A comparison restricted to adherers can mix those differences into what looks like a treatment effect.

The Label Problem: “ITT” Does Not Always Mean the Same Thing

The SPIRIT 2025 explanation notes that many analyses are described as “intention-to-treat” while applying different rules for handling missing outcome data or for excluding participants who deviated. It advises that labels such as “intention-to-treat” or “modified intention-to-treat” be avoided unless the protocol fully defines them (SPIRIT 2025, Item 27b).

Protocols posted on ClinicalTrials.gov show how definitions differ:

  • One protocol defines three populations. ITT is all randomized participants. Modified ITT is randomized participants who received at least one dose. Per-protocol is ITT participants without major protocol deviations. That trial planned the modified ITT analysis as primary and per-protocol as secondary (NCT04391309 protocol and analysis plan).
  • Another defines ITT as all randomized subjects and designates it as the primary efficacy population, with the per-protocol analysis described as supportive (NCT02164981 analysis plan).

The practical point: when a paper or summary says “ITT,” confirm who was actually counted. If the number analyzed is smaller than the number randomized, the report should explain who was excluded and why.

What the Evidence Says About Effect Sizes

The SPIRIT 2025 explanation cites a meta-epidemiological study of 322 comparisons from 310 randomized trials. That study found that analyses deviating from intention-to-treat produced larger intervention effect sizes than analyses applying the ITT principle (SPIRIT 2025, Item 27b). This is an average pattern across trials, not proof that every per-protocol result is inflated. It is a reason to check which analysis a headline number comes from.

Where to Find Each Result in a Trial Report or Registry

The CONSORT 2025 checklist for reporting completed trials includes several items that let readers see both views (CONSORT 2025 explanation and elaboration, BMJ 2025):

  • Item 21b: Who is included in each analysis, and in which group.
  • Item 21c: How missing data were handled.
  • Items 22a and 22b: For each group, the numbers randomized, the numbers who received the intended intervention, and the numbers analyzed for the primary outcome. They also cover losses and exclusions after randomization, with reasons. These usually appear in the participant flow diagram.
  • Item 24a: How the intervention and comparator were actually delivered, including whether participants adhered.
  • Item 26: For each outcome and group, the number included in the analysis and the number with available data at the outcome time point.
  • Item 28: Any other analyses, including sensitivity analyses, distinguishing prespecified from post hoc.

For trials covered by US federal results-reporting rules, the required results information includes participant flow, baseline characteristics, and outcome measures. Each outcome reports a Number of Participants Analyzed. When that number differs from the number assigned to a group, an Analysis Population Description must briefly explain why (42 CFR 11.48). On a ClinicalTrials.gov results page, comparing the number analyzed with the number who started is a quick way to see whether a result covers the full randomized group.

What Remains Uncertain

  • Better per-protocol methods exist, but they depend on assumptions. The original Coronary Drug Project analysis reported a 9.4-percentage-point five-year mortality difference between adherers and non-adherers in the placebo arm. A 2016 reanalysis using modern adjustment methods reduced that difference to 2.5 percentage points. The authors note that valid per-protocol estimates generally require adjusting for prognostic factors associated with adherence (Murray and Hernán, Clinical Trials 2016). Whether a given trial collected the data needed for that adjustment has to be judged trial by trial.
  • The “estimand” framework is newer. The CONSORT 2025 explanation describes the ICH E9(R1) estimands framework, which asks trials to state in advance how events after randomization, such as stopping treatment, will be handled. Including estimands in the CONSORT 2025 checklist did not reach consensus, though the authors expect the framework to become more widespread (CONSORT 2025 E&E, Box 1).
  • Missing data can affect both analyses. An ITT analysis with substantial missing outcomes depends on how those outcomes were handled, which is why CONSORT 2025 asks for it separately in Item 21c.

Your Next Practical Step When Reading a Trial

  1. Find the number randomized in each group, usually in the flow diagram or registry participant flow.
  2. Find the number analyzed for the primary outcome. If it is smaller, find out who was left out and why.
  3. Identify which analysis was prespecified as primary, ideally by checking the registered protocol or statistical analysis plan.
  4. Check adherence and dropout by group. Uneven dropout between arms deserves extra attention.
  5. Compare the ITT and per-protocol results if both are reported, and note whether the conclusion changes.
  6. Be cautious with any summary that cites only “completers” or adherent participants without showing the full randomized comparison.

For related guides on reading trial methods, browse the Study Design section. For how trial limits shape what results can and cannot tell you, see Safety & Limitations. You can learn more about this publication on our About page.

Important Limits of This Guide

This article explains how clinical trial results are analyzed and reported. It is educational, does not evaluate any specific product, treatment, or supplement, and is not medical advice. Understanding how a trial was analyzed does not tell you whether a treatment is right for you. Do not start, stop, or change any medication or supplement based on a study result alone; talk with a qualified healthcare professional who knows your health history.

Sources