Intention-to-Treat vs Per-Protocol Analysis

Intention-to-treat analyses every randomised participant in the group they were allocated to, whether or not they took the drug, completed the course or stayed in the study. Per-protocol restricts the analysis to those who followed the protocol as written. The distinction matters because randomisation balances known and unknown prognostic factors only at the moment of allocation, so once participants are removed on the basis of what happened afterwards the comparison becomes observational.

Regulatory practice now asks for the question before the analysis. The ICH E9 addendum on estimands, adopted in 2019, requires trials to specify in advance how intercurrent events such as treatment discontinuation and rescue medication are handled, precisely because a phrase like intention-to-treat had come to cover several incompatible approaches. A treatment-policy estimand asks what happens when a drug is prescribed, including in those who stop it; a hypothetical estimand asks what would happen if everyone continued.

Direction of bias is the practical point. In a superiority trial intention-to-treat is generally conservative, because non-adherence dilutes any real difference toward the null, so a positive result is hard to dismiss. In a non-inferiority trial the incentive reverses, since non-adherence and sloppy conduct make two treatments look more alike and therefore make non-inferiority easier to claim, which is why such trials must report both analyses and show they agree.

The misuse is a per-protocol result presented as the headline with the intention-to-treat number in a supplement, or absent. That is the same manoeuvre as reporting completers only, and it flatters a drug whose users stop for side effects or lack of benefit. The mirror-image error is treating intention-to-treat as the biological effect in someone who takes the drug faithfully, which it is not.