Should Both Intention-to-Treat and Per-Protocol Analyses Be Planned in Noninferiority Trials?

Should a Noninferiority Trial Analyze Everyone Randomized and Only Those Who Followed the Protocol?

You are preparing the statistical analysis plan for a noninferiority trial. One recommendation is to analyze participants according to their randomized treatment assignments. Another is to examine participants who adhered sufficiently to the protocol.

Which analysis should determine whether the new intervention is noninferior?

The answer is less straightforward than choosing the analysis that produces the smaller treatment difference. Intention-to-treat and per-protocol analyses address different methodological concerns, and neither is automatically free from bias. Their suitability depends on the treatment effect the study intends to estimate and the assumptions supporting each analysis.

Plan Both When They Provide Relevant, Complementary Evidence

How Intention-to-Treat and Per-Protocol Analyses Differ in Noninferiority Research

What Is Intention-to-Treat Analysis?

Under the intention-to-treat principle, randomized participants are analyzed according to the groups to which they were originally assigned, regardless of subsequent adherence, treatment switching, or discontinuation.

The principle helps preserve the comparison established by randomization. It avoids selectively reassigning or excluding participants merely because their post-randomization treatment experience differs from the original protocol.

However, the label intention-to-treat does not specify how missing outcome data are handled, nor does it fully define the treatment effect being estimated.

A complete analysis plan must distinguish retaining randomized assignment from the statistical methods used to address missing outcomes and intercurrent events.

For example, an analysis may retain participants in their randomized groups while making assumptions about outcomes after treatment discontinuation. Those assumptions affect its interpretation.

What Is Per-Protocol Analysis?

Per-protocol analysis generally evaluates outcomes among participants who satisfy specified protocol requirements, such as receiving an adequate amount of assigned treatment and avoiding major deviations.

Its purpose is often to examine treatment effects under conditions closer to the intended intervention regimen.

However, there is no single universally applicable definition of a per-protocol population. The criteria depend on the intervention, trial protocol, and scientific objective.

Some studies exclude participants with major deviations. Others use more sophisticated approaches to estimate effects under adherence rather than relying only on exclusions.

Because adherence and protocol deviations occur after randomization, simple per-protocol restriction may disrupt the comparability created by random assignment.

The term per-protocol should therefore be accompanied by a precise description of the population, treatment conditions, and statistical methods.

Why Are These Analyses Particularly Important in Noninferiority Trials?

In a superiority trial, nonadherence and treatment crossover may reduce the observed contrast between interventions, making a genuine difference harder to detect.

In a noninferiority trial, reduced treatment contrasts may sometimes make the interventions appear sufficiently similar to satisfy the noninferiority criterion.

This creates concern that an analysis retaining participants according to randomized assignment could support apparent noninferiority when the interventions would differ more substantially under adequate treatment receipt.

Nevertheless, this concern does not imply that intention-to-treat analysis always favors noninferiority.

The direction of bias depends on treatment effects, patterns of adherence, crossover, missing outcomes, and the estimand being targeted.

Likewise, a per-protocol analysis may introduce selection bias when participants who adhere differ prognostically from those who do not.

Understanding how poor adherence can distort noninferiority comparisons is therefore essential before deciding how either analysis should be interpreted.

How Do Intention-to-Treat and Per-Protocol Analyses Compare?

The comparison describes common analytical approaches. Specific methods and their validity depend on the estimand, trial design, and missing-data assumptions.

Why Does the Estimand Framework Change How These Analyses Are Planned?

ICH E9(R1) emphasizes that the treatment effect of interest should be defined before selecting the estimator or statistical analysis.

An estimand specifies the population, treatment conditions, outcome variable, handling of intercurrent events, and population-level summary.

Intercurrent events include events after treatment initiation that affect the interpretation or existence of measurements relevant to the research question. Examples may include treatment discontinuation, rescue medication, and crossover.

The estimand framework helps clarify whether a trial seeks to evaluate outcomes under the treatment strategies as assigned or under a hypothetical condition in which certain deviations did not occur.

These are different scientific questions, even when both are relevant to the same treatment decision.

Does Intention-to-Treat Analysis Always Estimate a Treatment-Policy Effect?

Not automatically. An analysis retaining randomized assignment can support a treatment-policy estimand when outcomes are collected and analyzed appropriately regardless of relevant intercurrent events.

However, if outcomes after discontinuation are missing or replaced using assumptions corresponding to a hypothetical scenario, the resulting analysis may not estimate a treatment-policy effect.

The phrase intention-to-treat is therefore insufficient to identify the estimand.

Researchers should explicitly describe how outcomes after treatment switching, discontinuation, and rescue treatment are handled.

Does Per-Protocol Analysis Estimate the Effect of Full Adherence?

Not necessarily.

A simple analysis restricted to participants who adhered sufficiently may compare selected groups with different prognostic characteristics.

For example, participants who experience adverse events or early treatment failure may be more likely to discontinue. Excluding them can systematically change the outcome distribution.

As a result, the difference among observed adherers may not equal the causal effect that would have occurred if everyone had adhered.

Estimating such an effect may require methods that address time-varying adherence and confounding, including appropriate causal inference techniques.

These approaches can be useful, but their validity depends on assumptions that may not be fully testable.

Why Is Agreement Between the Analyses Useful but Not Definitive?

When intention-to-treat and per-protocol analyses both support noninferiority, investigators may gain confidence that the conclusion is not solely an artifact of one particular analysis population.

However, agreement does not prove that either analysis is unbiased.

Both analyses may be affected by missing data, measurement error, poor assay sensitivity, or assumptions that do not hold.

Moreover, if the analyses target different estimands, disagreement does not necessarily mean that one is incorrect.

It may reflect genuinely different effects of treatment assignment and sustained treatment receipt.

Researchers should therefore interpret agreement or disagreement in relation to the questions each analysis was designed to answer.

What If the Analyses Reach Different Conclusions?

Suppose the intention-to-treat analysis supports noninferiority but the per-protocol analysis does not.

This pattern may raise concerns about adherence-related dilution, selection effects, or differences in the treatment effects being estimated.

The opposite pattern is also possible. A per-protocol analysis may support noninferiority while the intention-to-treat analysis remains inconclusive or suggests an unacceptable disadvantage.

Neither pattern should automatically be resolved by selecting the more favorable result.

Instead, investigators should examine the prespecified primary estimand, the magnitude and precision of estimates, protocol deviations, missingness, and the assumptions underlying each analysis.

When disagreement cannot be adequately explained, the strength of the overall noninferiority claim may need to be qualified.

Should Both Analyses Be Required to Demonstrate Noninferiority?

There is no universal rule that every noninferiority trial must obtain statistically significant noninferiority results from both an intention-to-treat and a per-protocol analysis.

Regulatory and methodological guidance has traditionally encouraged careful examination of both approaches because each can be vulnerable to different biases.

However, contemporary estimand-based planning places greater emphasis on selecting analyses that correspond to explicitly defined treatment effects.

Whether a supplementary analysis must satisfy a particular decision criterion depends on the protocol, applicable regulatory requirements, and the intended confirmatory interpretation.

The decision should be prespecified rather than invented after the analyses are complete.

How Do Missing Data Affect Both Approaches?

Missing outcome data create a problem distinct from, although often related to, treatment nonadherence.

A participant may discontinue treatment but remain available for outcome assessment. Conversely, a participant may adhere to treatment but have an unobserved outcome.

Researchers should avoid treating treatment discontinuation and missing data as the same event.

Both intention-to-treat and per-protocol analyses can be biased when missingness is handled using inappropriate assumptions.

For noninferiority trials, assumptions that make outcomes appear more similar can be especially consequential.

Prespecified sensitivity analyses should therefore examine whether the conclusion remains credible under plausible alternative missing-data assumptions.

How Do the Noninferiority Margin and Assay Sensitivity Affect Interpretation?

Neither analysis population can compensate for an unjustified noninferiority margin.

If the margin permits a clinically unacceptable disadvantage, obtaining favorable results from both analyses does not make that threshold appropriate.

Similarly, neither analysis can independently establish that the active comparator retained meaningful effectiveness in the trial.

The credibility of assay sensitivity remains essential when interpreting an active-controlled comparison.

The justification of the noninferiority margin, the estimand, and the statistical analysis strategy should therefore be developed together.

What Happens When Intention-to-Treat and Per-Protocol Analyses Disagree?

The example does not establish that one analysis is necessarily correct and the other incorrect. It shows why apparently contradictory findings require investigation rather than selective reporting.

Importantly, a failure to establish noninferiority in the supplementary analysis does not by itself prove that the new treatment is unacceptably inferior. The confidence interval may simply be insufficiently precise to exclude that possibility.

Common Errors When Using ITT and Per-Protocol Analyses

How to Plan ITT and Per-Protocol Analyses Without Choosing Results Retrospectively

Begin with the treatment effect your study intends to estimate. Define the population, interventions, outcome, intercurrent-event strategies, and population-level summary before selecting the analysis method.

Then determine which analysis best addresses that estimand and which supplementary analyses would provide meaningful information about robustness or alternative treatment effects.

In the statistical analysis plan, distinguish the primary confirmatory analysis from supplementary and sensitivity analyses. Define protocol deviations considered important, how adherence will be measured, and how outcomes after discontinuation or crossover will be handled.

Researchers should also explain whether an adherence-related analysis is intended to estimate a causal effect or merely describe outcomes in a selected subset.

When major protocol deviations or unreliable comparator effectiveness make the intended conclusion difficult to defend, researchers should consider whether the noninferiority objective remains appropriate.

Before Finalizing ITT and Per-Protocol Analyses

Frequently Asked Questions About ITT and Per-Protocol Analyses in Noninferiority Trials

Plan Analyses Around the Estimand, Not Around Which Result Looks Better

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.