Per-Protocol Analysis & LOCF
Per-protocol analyses only the patients who adhered to the protocol — which breaks randomisation; LOCF fills missing follow-up with each patient’s last recorded value — an outdated, biased fix.
Per-protocol (PP) includes only participants who completed the assigned treatment as specified. It can answer “does it work if actually taken?” but breaks randomisation — adherers differ systematically from non-adherers. LOCF (last observation carried forward) imputes each patient’s missing follow-up with their last recorded value, assuming a frozen trajectory.
(non-adherent / dropout)
(non-adherent / dropout)
Dropping non-adherers leaves self-selected, unbalanced arms — randomisation is broken.
A CONSORT-style flow where non-adherent and dropout patients are removed from each arm before analysis. Unequal, non-random exclusions (90 vs 40 here) leave two arms that are no longer the comparable groups randomisation created — the analysed 870 are a self-selected subset.
Whenever patients are dropped after randomisation for what happened next, suspect bias. Adherers tend to be healthier (the healthy-adherer effect), so PP often flatters the treatment. Read PP only as a secondary, supportive analysis — the primary result of a superiority trial should be ITT.
PP and LOCF can each manufacture a more favourable answer than the truth. PP matters chiefly in non-inferiority trials (where ITT is anti-conservative), but presenting it as the primary superiority result — or patching dropouts with LOCF — is a red flag for spin.
Per-protocol= adherers only → breaks randomisationITTstays the primary analysis of a superiority trialLOCF= outdated imputation; prefer mixed models / multiple imputation
Reporting both ITT and per-protocol — a well-conducted RCT pre-specifies ITT as primary and presents per-protocol alongside as a sensitivity analysis. When the two agree, confidence rises; when PP looks markedly better than ITT, that gap usually reflects the non-random loss of non-adherers, not a larger true effect. Beware the trial that headlines the per-protocol number while burying a null ITT result — or one that “handled” dropouts with LOCF. Both quietly reintroduce the selection bias randomisation was designed to remove.
- Per-protocol breaks randomisation — excluded non-adherers differ systematically from completers.
- LOCF bias — carrying a last value forward assumes a frozen trajectory and is no longer accepted practice.
- Selectively reporting whichever of ITT / PP looks best — outcome-analysis spin.
Quick check
Why isn’t per-protocol the primary analysis of a superiority trial?
Answer: Because excluding non-adherers breaks the randomised balance — adherers differ systematically from non-adherers (the healthy-adherer effect), so PP can exaggerate the treatment effect. ITT preserves randomisation and is the conservative, unbiased primary analysis.
Ready to build your plan? EMF Premium gives you all 40,000+ questions, 20 mocks and 1,215 OSCE stations from £29/month — or a one-off 3- or 6-month pass.