Clicky

Clinical Research Explained

Plain-language explanations of health research and the products built on it

Per-protocol and intention-to-treat analysis: Why they answer different research questions

Intention-to-treat (ITT) and per-protocol (PP) analyses answer two different questions about the same trial. An intention-to-treat analysis counts every randomized participant in the group they were assigned to, whether or not they followed the plan, so it estimates the effect of being assigned to a treatment. A per-protocol analysis keeps only participants who followed the protocol, so it tries to estimate the effect of actually receiving the treatment as specified. The CONSORT reporting guidance describes intention-to-treat as the widely recommended default for randomized trials because it preserves the benefit of randomization, and says a per-protocol analysis should be labeled a non-randomized, observational comparison.

This article is general education about research methods, not medical advice. It has not been peer reviewed or clinically reviewed. Sources were checked on October 9, 2026. See the Medical Disclaimer.

Start with the protocol: what “per-protocol” refers to

The National Institutes of Health explains that clinical trials follow a plan called a protocol, which is designed to answer specific research questions. The protocol describes who is eligible, which tests and treatments take place, how long the trial lasts, and what information is collected. “Per-protocol” is therefore a statement about adherence to that plan.

The same NIH page describes randomization as assigning treatments to participants by chance rather than by choice, to avoid bias in who receives which treatment. That matters here, because the two analyses treat randomization very differently.

What each analysis answers

  • Intention-to-treat: All randomized participants are analyzed in the group they were assigned to, including people who stopped early, switched, or never started. The question is, “What happened to people offered this treatment, given real-world adherence?”
  • Per-protocol: Only participants who met the protocol's adherence requirements are analyzed. The question is, “What happened to people who actually received the treatment as specified?”

A 2020 methods review by Tripepi and colleagues describes the two as complementary strategies: intention-to-treat assesses the effect of assigning a drug, and per-protocol assesses the effect of receiving the assigned treatment as specified in the protocol. The authors say both are essentially valid but have different scopes and interpretations depending on context.

Why the two numbers can diverge

The CONSORT 2010 explanatory document says random assignment, when properly done, balances both known and unknown prognostic factors between groups. In its words on intention-to-treat, including all randomized participants in their allocated groups is what fully preserves that benefit. Once participants are removed after randomization, the groups being compared are no longer the ones chance created, and CONSORT says any such exclusion compromises the randomization and may bias the results.

CONSORT also notes that people excluded after allocation are unlikely to be representative of everyone in the study. For example, someone may be unavailable for follow-up because their illness worsened or because of harms from treatment. At the same time, because intention-to-treat counts people who did not receive the intervention as allocated, CONSORT points out that a reader can use the number of non-recipients to judge how far the estimated efficacy might fall short of what would be seen under ideal circumstances. That gap is part of what a per-protocol analysis tries to capture.

CONSORT adds that a strict intention-to-treat analysis is often hard to achieve, for two reasons:

  • Missing outcomes. Some participants have no recorded final result. Analyzing only those with results (a “complete case” analysis) means the analysis is no longer strictly intention-to-treat, and CONSORT says bias may be introduced if being lost to follow-up is related to how a person responded to treatment. Filling in missing results by estimation (imputation) keeps everyone in the analysis but requires strong assumptions that can be hard to justify.
  • Non-adherence to the protocol. Some participants do not take all the intended treatment, receive a different treatment, or turn out not to meet the entry criteria. The simple intention-to-treat approach is to include them anyway. Excluding them for these reasons is incompatible with intention-to-treat.

A worked comparison (hypothetical numbers)

Suppose a reader sees a summary that says: “Among participants who followed the plan, 60% responded to Intervention A, compared with 40% on the comparator.” The numbers below are invented to show how that sentence can differ from the full trial result. They do not come from a real study.

  • 400 people were randomized: 200 to Intervention A and 200 to the comparator.
  • Arm A: 150 followed the protocol and 50 stopped early. Of the 150 adherent participants, 90 responded. Of the 50 who stopped, 10 responded.
  • Comparator arm: 190 followed the protocol and 10 did not. Of the 190 adherent participants, 76 responded. Of the 10 who did not, 4 responded.

Per-protocol result: 90 of 150 (60%) versus 76 of 190 (40%), a 20-percentage-point difference. This is the sentence in the summary.

Intention-to-treat result: 100 of 200 (50%) versus 80 of 200 (40%), a 10-percentage-point difference.

In this invented example the gap shrinks because 50 people in Arm A stopped early, and they responded less often (20%) than the people who stayed (60%). If people who stop early differ in some systematic way, such as stopping because they felt no benefit, then dropping them leaves a more favorable subgroup. The per-protocol comparison is then between adherent people in each arm, and adherence itself may be linked to the outcome. That is a different comparison from the one randomization protected. Notice also that the two arms lose different numbers of people (50 versus 10), which is the kind of imbalance CONSORT says deserves attention.

Neither number tells a reader what will happen to them. Both describe groups of participants in one trial.

How to read a result that cites one or the other

  1. Find the numbers behind each analysis. CONSORT Item 16 asks authors to report, for each analysis, how many participants in each group were included and whether the analysis used the original assigned groups. It also says results should be shown as fractions or event rates, not only as summary measures such as relative risks, so you can see whether anyone randomized was left out. Our guide to reading a CONSORT participant-flow diagram shows where to look.
  2. Check which analysis the trial named as primary. A planned analysis is easier to trust than one chosen after seeing the data. See our explainer on primary and secondary outcomes, and our guide to protocol amendments if the plan changed.
  3. Do not take the label at face value. CONSORT cites a review of 119 trial reports that said all participants were analyzed as originally assigned. In 15 of them (13%), patients were excluded or not all were analyzed as allocated.
  4. Be cautious with “modified intention-to-treat.” CONSORT says this term is widely used for analyses that exclude participants who did not adhere adequately, and that neither it nor “intention-to-treat” reliably tells you who was included. For that reason, CONSORT 2010 asks authors to describe exactly who was included in each analysis.
  5. Look at the reason for any exclusion. CONSORT says that calling something a “protocol deviation” is not enough to justify excluding participants after randomization. The nature of the deviation and the exact reason should be reported.
  6. Compare the two when both are reported. A small gap suggests the conclusion does not hinge on who adhered. A large gap is a prompt to ask why people left and whether they differed from those who stayed.

Limits of both approaches

  • An “intention-to-treat” label does not by itself say how missing outcomes were handled. The methods section should say.
  • Intention-to-treat estimates can be diluted by non-adherence, so they may not show what a treatment does when taken as directed.
  • Per-protocol estimates can be distorted if adherence is related to prognosis or outcome. Methodologists studying this problem, such as Miguel Hernán in a 2020 colloquium abstract, state that estimating the per-protocol effect validly generally requires adjusting for factors linked to adherence and loss to follow-up, both before and after randomization. Simply dropping non-adherent participants does not do that.
  • This article draws on the CONSORT 2010 explanatory document and two short methods abstracts. It does not cover regulatory guidance or every statistical method used to handle non-adherence.
  • Neither analysis diagnoses a condition, predicts an individual's outcome, or guarantees a result. Questions about a personal health decision belong with a qualified clinician.

Takeaway

Intention-to-treat asks what happens to people assigned to a treatment, and per-protocol asks what happens to people who follow the plan. A trustworthy report states which analysis it used, shows how many people each analysis includes, and explains any exclusions after randomization. For more on reading trials this way, see our How to Read Health Research hub.

Sources

Related guides

Related guides