An intention-to-treat analysis counts every participant in the group they were randomised to, regardless of what they actually did.
An intention-to-treat analysis counts every participant in the group they were randomised to, regardless of what they actually did. People who stopped, switched or never started are still analysed where they were assigned.
Including people who did not take a treatment appears to dilute its effect, and it does. The alternative reintroduces selection: people who complete a course differ systematically from people who do not, and comparing completers destroys the randomisation.
Restricting to those who followed the protocol estimates the effect under ideal adherence. It is informative and no longer protected by randomisation, so it belongs as a secondary analysis rather than the headline.
A trial presenting only the completers' result has chosen the analysis most likely to favour the treatment. Where the intention-to-treat figure is absent entirely, that absence is the most important thing on the page.
Participants lost to follow-up have no outcome to analyse. Methods for handling this range from conservative assumptions to statistical imputation, and the choice can change the conclusion, so good trials report a sensitivity analysis across several.
Intention-to-treat estimates what happens when a treatment is offered in practice, including people who will not take it. That is usually the more useful quantity for a decision, and it is systematically smaller than the ideal-adherence figure.
Analyses described as modified exclude some randomised participants on stated criteria. Sometimes reasonable, sometimes a per-protocol analysis with a better name, and the exclusions should be specified precisely enough to check.
Whether an intention-to-treat result is reported, how many were missing, and whether the conclusion survives different assumptions about them.
When the question is whether a treatment is no worse than an existing one, dilution from non-adherence makes two treatments look more similar, which favours the conclusion the trial is testing for. Intention-to-treat is therefore the conservative choice in superiority trials and the permissive one here, which is why non-inferiority trials report both.
Decisions about imputation made after seeing which data are missing can move a result substantially. A protocol naming the primary approach, with sensitivity analyses across plausible alternatives, is the standard; choosing afterwards is a known route to a favourable answer.
A trial presenting only the per-protocol result has selected the analysis most favourable to the treatment. Where the intention-to-treat figure is absent entirely, that absence is the most important thing on the page.
Intention-to-treat estimates what happens when something is offered in practice, including to people who will not persist with it. That is generally the quantity a decision needs, and it is systematically smaller than the ideal-adherence figure.
Participants who leave a trial still need a value in the analysis, and the method chosen for supplying it can move the result substantially. Trials reporting several approaches and showing the conclusion survives all of them are doing the thing that makes an estimate trustworthy.