Per-Protocol Versus Intention-to-Treat: When Each Answers the Real Question

Intention-to-treat and per-protocol analyses answer two different questions, and the confusion between them explains a lot of arguments about trial results. Intention-to-treat keeps every randomized participant in the group they were assigned to, regardless of what they actually did, which preserves the balance that randomization bought and estimates the effect of a treatment policy in the real world. Per-protocol restricts the analysis to people who followed the protocol closely, which sounds more honest but quietly breaks randomization and can reintroduce the very bias the trial was built to remove. For a superiority trial, intention-to-treat is usually the primary analysis for exactly that reason. This is a methods article, not medical advice; for anything about your own care, talk with a clinician who knows your history.

What does intention-to-treat actually mean

The rule is almost aggressively simple: analyze people in the group they were randomized to, even if they never took a single dose, stopped halfway, or crossed to the other arm. Once randomized, always analyzed. It feels wrong the first time you meet it. Why should a treatment get credit for someone who refused it?

The answer is that randomization is the only feature of a trial that makes the two groups comparable on everything, measured and unmeasured, before treatment starts. That balance is fragile. The moment you remove people based on what happened after randomization, you are sorting on the future, and the groups you compare are no longer the groups randomization created. Intention-to-treat protects the one thing that separates a randomized trial from a well-dressed observational study.

There is a second, quieter reason. In the clinic, patients also skip doses, stop early, and change their minds. A treatment that works beautifully but that half of people cannot tolerate is not as useful as its biology suggests. Intention-to-treat folds real-world adherence into the estimate, so it answers the question a health system actually faces: if we adopt this policy, what happens on average to the people we offer it to?

What is per-protocol trying to estimate

Per-protocol asks a narrower and, on its face, reasonable question: among people who took the treatment as intended, how well did it work? This is closer to a biological efficacy question, the effect of the drug when the drug is genuinely present in the body. Clinicians care about this. If I want to know whether a molecule does what its mechanism promises, adherence noise is exactly what I want to strip away.

The trouble is in how you strip it. To build a per-protocol population you exclude people who did not adhere, and adherence is not random. People who stay on a treatment often differ from those who stop, and they differ in ways that also affect outcomes. They may be healthier, better supported, or simply spared the side effects that push others to quit. When you keep the adherent and drop the rest, you are no longer comparing like with like. You have re-created a selected comparison inside a randomized trial, and the protection is gone.

A common trap is the assumption that non-adherence is a nuisance sprinkled evenly across both arms. It rarely is. If the active drug causes more side effects, the people who drop out of the treatment arm are systematically different from the people who drop out of the placebo arm, and the survivors you compare are mismatched in a direction you cannot see.

Why does per-protocol tend to flatter the treatment

A classic and humbling finding in the trial-methodology literature belongs in every appraisal course. Among participants assigned to a placebo, those who took it faithfully had markedly better survival than those who did not. Nobody thinks the placebo was working. Adherence was a marker for a whole cluster of healthy behaviors and circumstances. The lesson generalizes: adherent people do better for reasons that have nothing to do with the pill, so any analysis that conditions on adherence inherits that advantage.

Apply that to a real treatment and per-protocol will usually push the estimate away from the null, making the treatment look stronger than the policy effect would suggest. Sometimes that larger number is closer to the true biological efficacy. Often it is closer to wishful thinking. You cannot tell which from the per-protocol estimate alone, and that is precisely the problem.

When is per-protocol the more honest choice

There is one important reversal. In a non-inferiority trial, where you are trying to show a new treatment is not meaningfully worse than an established one, intention-to-treat becomes the lenient analysis rather than the strict one. Non-adherence blurs the difference between arms, dragging both toward each other, which makes two treatments look more alike. That blurring makes it easier to declare non-inferiority. Here intention-to-treat can flatter the weaker conclusion, so careful trialists report both analyses and lean on per-protocol to keep themselves honest. The direction of caution flips with the question being asked.

How should a careful reader read the gap between them

Look for both numbers, and treat the space between them as information rather than an inconvenience. A well-reported trial gives intention-to-treat as primary and per-protocol as a sensitivity analysis, and the two should tell a coherent story.

When they agree, your confidence rises. The result is not an artifact of who happened to stay in the study. When per-protocol is substantially larger than intention-to-treat, resist the urge to quote the bigger number. That gap is usually telling you that dropout was informative, or that the treatment works well for the people who can stay on it but not for everyone offered it. That is a real and useful message, but it is a message about selection, not a license to upgrade the effect.

Three quick habits help. Check how non-adherence and crossover were handled and how many people were affected, because a per-protocol analysis that discards a large fraction of participants deserves suspicion. Ask whether dropout differed between arms, since asymmetry is where bias hides. And match the analysis to the claim: for the practical question of adopting a treatment, intention-to-treat is the honest anchor, and per-protocol is a lens you hold up beside it, never instead of it.