03 · What You Need to Know
How Nonadherence Can Distort a Noninferiority Comparison
What Counts as Poor Adherence in a Noninferiority Trial?
Adherence refers to the extent to which participants follow the treatment regimen or intervention procedures specified for them. Its precise definition depends on the intervention and the scientific question.
For medication trials, nonadherence may involve missed doses, incorrect dosing, premature discontinuation, or failure to begin treatment. In rehabilitation or behavioral intervention trials, it may involve incomplete sessions, inconsistent participation, or inadequate exposure to the intervention.
Treatment crossover occurs when participants receive the alternative intervention, while contamination occurs when treatment elements intended for one group are received by another.
These events are not interchangeable. A participant who stops treatment because of an adverse event presents a different interpretive problem from someone who misses appointments for unrelated reasons.
Researchers should define adherence operationally before the trial and distinguish the amount of intervention received from the reasons for deviations.
Why Is Treatment-Effect Dilution Particularly Concerning for Noninferiority?
In a conventional superiority trial, treatment-effect dilution may reduce the ability to detect a genuine advantage. In a noninferiority trial, the same phenomenon may sometimes make the interventions appear sufficiently similar to satisfy the noninferiority criterion.
Consider an effective standard treatment and a less effective new treatment. If many participants assigned to standard care do not receive an adequate course, the observed benefit of standard care may decline.
The difference between groups can then become smaller than the difference that would have occurred with adequate delivery of both interventions.
When the lower confidence bound remains above the prespecified noninferiority margin, researchers may conclude that noninferiority has been demonstrated.
Yet the apparent similarity could partly reflect inadequate implementation of the comparator rather than adequate effectiveness of the new intervention.
This is why assay sensitivity is important: the trial must have a credible ability to distinguish treatments with genuinely different effects.
Does Poor Adherence Always Bias Results Toward Noninferiority?
No. The common warning that nonadherence favors noninferiority describes a possible mechanism, not a universal statistical law.
If nonadherence occurs mainly in the new-treatment group, it may increase the apparent disadvantage of that intervention. If participants assigned to an inferior intervention switch to effective standard care, the observed groups may become more similar. Other patterns may produce different effects.
Nonadherence can also be associated with prognosis. Participants experiencing early treatment failure, adverse effects, or worsening symptoms may be more likely to discontinue treatment.
Consequently, the relationship between adherence and outcomes may involve selection, confounding, and time-dependent processes.
The direction of bias cannot reliably be inferred from the overall percentage of participants who adhered. Investigators need information about the timing, reasons, and distribution of deviations.
How Do Crossover and Contamination Affect the Comparison?
Crossover and contamination can reduce the contrast between randomized treatment strategies.
Suppose participants assigned to the new intervention receive substantial amounts of standard care after experiencing poor responses. Their eventual outcomes may improve because of the additional treatment.
An analysis retaining participants in their originally assigned groups can then estimate a smaller difference between treatment strategies than would be observed if each intervention were administered without crossover.
Whether that result constitutes bias depends partly on the treatment effect the study intends to estimate.
If the research question concerns the consequences of initially assigning a treatment strategy that permits subsequent switching, the observed crossover may be part of the relevant strategy.
If the objective concerns outcomes under sustained adherence to the assigned intervention, crossover may prevent the unadjusted comparison from answering that question.
The distinction should be addressed through a clearly defined estimand rather than decided retrospectively.
Why Does the Estimand Matter?
An estimand describes the treatment effect the trial intends to estimate. Under the ICH E9(R1) framework, its attributes include the population, treatment conditions, outcome variable, handling of intercurrent events, and population-level summary.
Intercurrent events are events occurring after treatment initiation that affect either the interpretation or existence of measurements relevant to the research question. Treatment discontinuation, rescue medication, and crossover may qualify.
Different estimands can lead to different interpretations of nonadherence.
Treatment-policy strategy
Evaluates outcomes regardless of specified intercurrent events, such as discontinuation or switching, when those events are part of the intended treatment comparison.
Hypothetical strategy
Addresses what outcomes would have been under a specified hypothetical condition, such as if treatment discontinuation had not occurred. Estimation requires assumptions about unobserved outcomes.
Other estimand strategies may also be appropriate. The important requirement is that the treatment effect, adherence expectations, and analysis approach correspond to the actual decision being investigated.
A treatment-policy effect is not inherently inferior to a hypothetical adherence effect. They answer different questions.
Why Does Intention-to-Treat Analysis Not Automatically Resolve the Problem?
Intention-to-treat analysis generally preserves randomized treatment assignment in the comparison. This protects an important feature of randomization and can support estimation of the effect of assignment to treatment strategies.
However, retaining participants in their assigned groups does not ensure that the estimated effect corresponds to sustained receipt of the intended interventions.
When substantial nonadherence reduces treatment contrasts, an intention-to-treat comparison may sometimes favor apparent noninferiority relative to an estimand focused on adequate treatment receipt.
That does not mean intention-to-treat analysis is intrinsically biased or inappropriate. Its suitability depends on the estimand and the assumptions required for the analysis.
Similarly, excluding nonadherent participants does not automatically produce a more valid estimate.
The methodological implications of intention-to-treat and per-protocol analyses in noninferiority trials require careful attention to their distinct assumptions.
Why Can a Simple Per-Protocol Analysis Introduce New Bias?
A per-protocol analysis attempts to evaluate outcomes among participants who satisfy specified protocol requirements. However, adherence occurs after randomization and may be related to prognosis.
For example, participants who tolerate treatment well may be more likely to complete it, while those with adverse effects or early deterioration may discontinue.
Restricting analysis to adherent participants can therefore create groups that are no longer comparable in the way randomization originally intended.
Even when the proportions of adherent participants are similar, their prognostic characteristics may differ.
More sophisticated methods for estimating effects under adherence, including appropriate causal inference approaches, may sometimes be needed. Such methods depend on additional assumptions and sufficiently detailed data.
Researchers should not regard per-protocol analysis as a bias-free correction for nonadherence.
How Can Poor Adherence Undermine the Established Comparator's Effectiveness?
Active-controlled noninferiority trials often rely on historical evidence that the established comparator is effective relative to placebo or another appropriate reference.
That evidence supports the expectation that the comparator will retain meaningful effectiveness in the new trial.
If comparator adherence is substantially poorer than in the historical studies, its observed effect may be diminished.
This creates uncertainty about whether the comparator would retain the treatment benefit assumed when the noninferiority margin was justified.
The problem is especially consequential when the new treatment's apparent noninferiority depends on the comparator maintaining a substantial historical benefit.
Thus, adherence is not merely an operational quality indicator. It can affect the assumptions supporting the scientific interpretation of the entire trial.
What Should Researchers Measure About Adherence?
Adherence measurement should match the intervention and the mechanisms through which deviations could affect outcomes.
Medication studies may use dispensing records, electronic monitoring, pill counts, treatment diaries, or other validated approaches. Each method has limitations. For example, dispensing a medication does not establish that it was consumed.
For behavioral interventions, researchers may record attendance, duration, participation, and intervention fidelity.
Where possible, adherence data should capture timing and intensity rather than relying only on a final binary classification.
Researchers should also document discontinuation reasons, rescue treatments, crossover, and relevant co-interventions.
These records help distinguish implementation failures from genuine differences in intervention performance.
How Should Adherence Problems Be Addressed in the Analysis?
The analysis plan should be developed before examining unblinded comparative outcomes.
Researchers should define the primary estimand, specify how intercurrent events will be handled, and identify sensitivity analyses that assess the robustness of conclusions to plausible assumptions.
Depending on the research question and available data, supplementary analyses may examine adherence-adjusted effects or outcomes under alternative assumptions.
However, post hoc exclusion of participants who did not adhere is not a reliable general solution.
The final report should describe adherence patterns by randomized group and explain how deviations may have affected the estimated treatment contrast.
When major deviations make the intended effect difficult to estimate, uncertainty should be acknowledged rather than concealed behind a statistically favorable noninferiority result.
