Ask Mallard

Intention-to-treat and per-protocol

Intention-to-treat analyses every randomized participant in the group they were assigned to, regardless of what happened afterwards.

What question does this help answer?

What happens if we ADOPT this policy, with the adherence we will actually get. That is a different question from what happens if everyone takes the treatment as directed, and both are legitimate.

Why it is the default

Randomization is what makes the groups comparable, and it is the only thing that makes them comparable in respect of unmeasured factors. Any exclusion made after randomization can undo that, because the reason someone stopped, switched or was found ineligible may be related to their prognosis. Intention-to-treat preserves the comparison the design created.

The two estimands

  • The INTENTION-TO-TREAT effect: the effect of being assigned to the treatment. Directly relevant to a policy decision, diluted by non-adherence.
  • The PER-PROTOCOL effect: the effect of actually receiving the treatment as specified. Relevant to an individual patient's decision, and not obtainable by simply dropping non-adherers.
  • Modified intention-to-treat: a term covering many different exclusions, some defensible and some not, so it means nothing without a precise definition.
  • As-treated: analysing by what people received, which abandons randomization entirely.

Who each analysis includes

Who each analysis includes after randomization. People assigned to A either take A or stop; people assigned to B either take B or switch to A. Intention-to-treat analyses everyone by assignment. A naive per-protocol analysis keeps only those who adhered. As-treated regroups people by what they actually received.After randomization: three analyses, three different groups of peopleRandomizedAssigned to AAssigned to BTook AStopped ATook BSwitched to AIntention-to-treat: everyone, by assignmentNaive per-protocol: only the adherent, byassignmentAs-treated: grouped by what was receivedAdherence happens after randomization, so dropping or regrouping people by it turns a randomizedcomparison into an observational one.
Conceptual diagram. It shows the structure of the idea, not data from any study.

Estimating a per-protocol effect properly

It requires adjustment for the post-randomization factors that predict adherence, using methods built for time-varying confounding such as inverse probability weighting or g-estimation, and it rests on assumptions the intention-to-treat analysis does not need. Instrumental variable methods using assignment as the instrument give a related quantity for the subgroup who would comply. All of this should be planned in advance, with the adherence data collected to support it.

Missing outcomes are the harder problem

Intention-to-treat requires an outcome for everyone randomized, and in practice some are missing. A complete-case analysis is not intention-to-treat, whatever it is labelled. The missingness mechanism must be named and handled, with a sensitivity analysis over plausible departures.

What to watch out for

Excluding people who did not adhere breaks randomization: adherence is a post-randomization behaviour, and the adherent patients in one arm are not comparable to the adherent patients in the other. A naive per-protocol analysis is therefore an observational comparison wearing a trial's clothes. And the frequently repeated claim that intention-to-treat is always conservative is false in a noninferiority trial, where blurring the arms works in favour of the conclusion being sought.

Clinical example

Question: Does a smoking cessation programme reduce relapse? Analysis: Intention-to-treat as the primary analysis including everyone randomized, with a prespecified per-protocol effect estimated by inverse probability weighting rather than by deleting the participants who stopped attending.

Recent studies that used this
Searching PubMed…
Part of the learning path Design a randomized trial (step 5 of 7).
Related Learn pages
Sources & further reading
  1. Intention-to-treat concept: A review (opens in a new tab) · Perspect Clin Res · doi 10.4103/2229-3485.83221
  2. Per-Protocol Analyses of Pragmatic Trials (opens in a new tab) · N Engl J Med · doi 10.1056/NEJMsm1605385
Last source review: 2026-09-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.