Per Protocol Analysis

Per protocol analysis is a method of analyzing data from a clinical trial that includes only those participants who completed the study in full compliance with the protocol. Participants who deviated from the protocol, including those who dropped out, switched treatments, or failed to adhere to the prescribed regimen, are excluded from the analysis.

Per protocol analysis aims to estimate the treatment effect under ideal conditions, answering the question: what happens when patients actually receive and adhere to the treatment as planned? This is a scientifically meaningful question, but per protocol analysis answers it only under a strong assumption: that the participants excluded from each arm do not differ in prognosis, so that removing them separates the effect of the intervention from the effects of nonadherence, crossover, and loss to follow up. However, by excluding noncompliant participants, per protocol analysis sacrifices the protection against confounding that randomization provides, potentially introducing bias. The comparison between per protocol results and intention to treat results is itself informative: when they agree, confidence in the treatment effect is strengthened; when they diverge, the reasons for the discrepancy warrant investigation.

In a per protocol analysis, the researcher defines criteria for protocol compliance before analyzing the data. These criteria specify what constitutes adequate adherence (for example, taking at least 80% of prescribed doses), acceptable follow up (completing all scheduled assessments), and protocol compliance (no violations of inclusion or exclusion criteria discovered after randomization).

Participants who meet all compliance criteria form the per protocol population, and the analysis is conducted on this subset. Those who deviate from the protocol in any defined way are excluded, regardless of their treatment assignment. This means that the groups being compared are no longer defined solely by randomization but also by post randomization behavior, which is influenced by treatment effects, side effects, patient characteristics, and disease severity.

This selection on post randomization factors is the fundamental limitation of per protocol analysis. Participants who adhere to treatment may be systematically different from those who do not. Adherent patients in the active treatment group may be healthier, more motivated, or experiencing fewer side effects than nonadherent patients. If the same is not true in the control group (where side effects may not drive nonadherence), the per protocol comparison is biased. This is sometimes called the "healthy adherer" effect, where adherence itself is a marker for better health behaviors and outcomes, independent of the treatment.

Regulatory agencies and clinical trial guidelines generally recommend intention to treat analysis as the primary analysis and per protocol analysis as a secondary or sensitivity analysis. For superiority trials, intention to treat analysis is considered more conservative and is preferred. For noninferiority and equivalence trials, however, intention to treat analysis is not necessarily conservative, because nonadherence tends to make treatments look more similar, potentially supporting a false conclusion of noninferiority. Per protocol analysis is not reliably conservative either, since its exclusions can bias it in either direction, so ICH E9 recommends planning both analyses, and a noninferiority conclusion is more convincing when both support it.

Include per protocol analysis as a secondary analysis alongside the primary intention to treat analysis, particularly in explanatory trials aimed at understanding the biological effect of a treatment. In noninferiority and equivalence trials, report both intention to treat and per protocol results, as ICH E9, the statistical guideline adopted by the FDA and EMA, recommends planning both. Always prespecify the definition of the per protocol population in the study protocol and statistical analysis plan.

A critical mistake is presenting per protocol analysis as the primary result of a superiority trial, which can overestimate treatment effects by excluding patients who may have discontinued treatment due to lack of efficacy or adverse events. Another error is defining the per protocol population post hoc (after seeing the data), which allows selection of criteria that favor a particular result. Failing to compare baseline characteristics between the per protocol population and the excluded participants is a further oversight, as systematic differences between these groups signal potential bias. Some researchers also fail to recognize that per protocol analysis is particularly important in noninferiority trials, where the intention to treat analysis may be anticonservative.

A randomized controlled trial compares a new oral anticoagulant to warfarin for stroke prevention in atrial fibrillation. Of 10,000 randomized patients (5,000 per arm), 800 in the new drug arm and 600 in the warfarin arm discontinue treatment before the end of the study. The intention to treat analysis includes all 10,000 patients and finds a hazard ratio of 0.85 (95% CI: 0.72 to 1.00, p = 0.05), a borderline significant result. The per protocol analysis includes only the 8,600 patients who completed the study on their assigned treatment and finds a hazard ratio of 0.78 (95% CI: 0.65 to 0.94, p = 0.008), a clearly significant result. The difference between the two analyses is explored: the higher dropout rate in the new drug arm was primarily due to gastrointestinal side effects, and the patients who dropped out had higher baseline stroke risk. Because more of these higher risk patients were removed from the new drug arm (800) than from the warfarin arm (600), the per protocol comparison may favor the new drug for reasons unrelated to its effect, so the intention to treat result remains the primary analysis and the per protocol result is read as a sensitivity analysis.