Choosing intention-to-treat or per-protocol
Intention-to-treat asks about assignment to treatment. Per-protocol asks about following a treatment strategy; dropping non-adherers does not estimate that effect.
Fit
When it matters, and when it does not
It matters when
- Your trial must distinguish the effect of offering treatment from the effect of following it.
- Participants may stop, switch or never start treatment, so the intended effect must be specified in advance.
It matters less when
- The study was not randomized. Specify its causal question and address confounding rather than invoking intention-to-treat. Read: Confounding
- The aim is to choose a regression model. Define the estimand first, then choose the method for the outcome and design. Read: The estimand
Before you start
What to check in your study
- Who counts as randomizedassignmentEveryone allocated, including people later found ineligible, unless a prespecified rule applied without knowing the arm says otherwise.
- Events after randomization that affect the outcome or its meaningtargetSuch as stopping treatment, switching arms or rescue medication. The estimands framework, ICH E9(R1), asks how each is handled.
- How missing outcomes will be handledIntention-to-treat needs an outcome for everyone. Name the missing-data assumption and plan a sensitivity analysis around it.
- Adherence data, if a per-protocol effect is wantedRecord adherence and what predicts it over time. Without them, a per-protocol effect cannot be estimated properly.
- Any modified intention-to-treat rule, defined exactlyThe term covers many different exclusions. State each one, why it is made, and that it was fixed before unblinding.
- Assignment was random and concealed.How to check: Intention-to-treat protects a randomized comparison. It cannot create one where allocation was predictable or subverted.
- Outcomes are known for everyone randomized, or missing ones are handled under a stated assumption.How to check: Report how many outcomes are missing in each arm, and why.
- People who stop or switch treatment are still followed.How to check: Keep collecting outcomes after people stop. Ending follow-up when treatment stops turns stopping into missing data.
Plain language
How it works
Intention-to-treat keeps each participant in the assigned arm, regardless of adherence. It addresses the effect of assignment with the adherence observed in that trial. A per-protocol effect asks what would happen if everyone followed the specified strategy. Dropping non-adherers does not estimate it. Factors that predict adherence may also predict outcomes.
Keep collecting outcomes after treatment stops. Missing outcomes still require assumptions under either estimand. A per-protocol effect may need methods for adherence and confounding over time; use the linked method articles. As-treated groups abandon randomization. An instrumental-variable effect for compliers is a different target again. Report these questions separately.
Worked example
A worked example
- Setting
- A trial assigns 600 smokers, 300 to a six-week quitting program and 300 to brief advice. Everyone's smoking status at 6 months is recorded.
- Question
- Does offering the program help people quit?
- Approach
- Compare participants by assigned arm to estimate the effect of offering the program. Treat the effect of full adherence as a separate question.
- Result
- By 6 months, 90 of 300 offered the program had quit, against 60 of 300 given brief advice. That is a risk difference of 10 per 100 and a risk ratio of 1.5.
- What it means
- Offering the program gives 10 more quitters per 100. Only 180 attended most sessions, and 72 of them quit, 40 in 100. Comparing them with brief advice is not randomized, because attenders may be more motivated.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Excluding people who did not adhere breaks randomization: adherence is a post-randomization behavior, 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 favor of the conclusion being sought.
- Calling a complete-case analysis intention-to-treat.Instead: Intention-to-treat needs everyone's outcome. Handle missing outcomes under a stated assumption, with a sensitivity analysis.
- Reporting modified intention-to-treat without saying who was left out.Instead: Define each exclusion, show it cannot depend on the assigned arm, and say when it was decided.
- Ending follow-up when someone stops treatment.Instead: Keep following them. Their later outcomes are part of the intention-to-treat effect.
- Deciding which effect matters after seeing the results.Instead: Name the estimand in the protocol, including how each event after randomization is handled.
Inside Mallard
How Mallard uses this
Plans do not link to this concept yet. It remains a step in Learn's randomized-trial path, after sample size and before ANCOVA.
concept.choosing_itt_per_protocol, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Intention-to-treat concept: A review (opens in a new tab) · Perspect Clin Res · doi 10.4103/2229-3485.83221
- Per-Protocol Analyses of Pragmatic Trials (opens in a new tab) · N Engl J Med · doi 10.1056/NEJMsm1605385
- The estimands framework: a primer on the ICH E9(R1) addendum (opens in a new tab) · BMJ · doi 10.1136/bmj-2023-076316