03 · What You Need to Know
Why Intervention Adherence Matters for Causal Comparisons
What Is Intervention Nonadherence?
Intervention nonadherence occurs when participants do not follow the treatment or intervention conditions specified by the study protocol or assigned treatment strategy.
Nonadherence can involve never initiating an assigned intervention, receiving only part of it, discontinuing it prematurely, using it inconsistently, or switching to another treatment.
For example, students assigned to complete ten weekly learning modules may complete only three. Patients assigned to a medication regimen may miss doses or discontinue treatment. Employees assigned to professional development may attend only selected sessions.
The relevant definition depends on the intervention. A single missed activity may have little substantive importance in one study but represent a meaningful departure from treatment in another.
Researchers should therefore define adherence operationally rather than assuming that participation is simply complete or incomplete.
Why Is Nonadherence Not Automatically a Failure of Randomization?
Random assignment determines which intervention participants are offered or assigned. It does not guarantee that every participant will receive or follow the intervention.
When randomization is implemented correctly, assignment is independent of baseline potential outcomes in expectation. Participants may subsequently behave differently, but those behaviors do not retroactively change the original assignment mechanism.
Consequently, a comparison of outcomes according to original randomized assignment can retain a causal interpretation for the effect of assignment, provided outcome ascertainment and other relevant assumptions are adequate.
However, the effect of being assigned to an intervention is not necessarily identical to the effect of actually receiving or adhering to it.
That distinction is central to understanding when nonadherence represents a threat to the causal comparison supporting the study's conclusion.
What Is the Difference Between Assignment, Receipt, and Adherence?
Researchers should distinguish the intervention assigned to participants from what they actually receive and how closely their behavior follows the intended strategy.
Treatment assignment
The intervention or comparison condition allocated to a participant, often through randomization.
Treatment receipt
The intervention components a participant actually receives or accesses.
Treatment adherence
The extent to which a participant follows the specified intervention strategy over the relevant period.
A student may be assigned to a learning platform, receive access credentials, and still complete none of the required activities. Assignment occurred, but meaningful intervention engagement did not.
Conversely, a student assigned to conventional instruction may obtain intervention materials through classmates, creating exposure outside the assigned condition.
The latter situation may also involve contamination between study groups.
Why Does the Causal Estimand Matter?
An estimand is the precisely defined quantity that a study seeks to estimate. In intervention research, it specifies the population, treatment conditions, outcome, and relevant approach to events occurring after assignment.
Nonadherence matters differently depending on the estimand.
If researchers want to know the effect of offering an educational program under realistic implementation conditions, incomplete uptake may be part of the policy-relevant result.
If they want to know the effect of sustained participation in the complete program, the effect of assignment may not answer that question.
Neither estimand is universally superior. They answer different research questions.
The International Council for Harmonisation's E9(R1) addendum emphasizes the importance of defining treatment effects and handling intercurrent events, including treatment discontinuation, consistently with the scientific question.
What Is Intention-to-Treat Analysis?
Intention-to-treat analysis generally compares participants according to their original randomized assignments, regardless of subsequent treatment receipt or adherence.
Its principal advantage is preserving the randomized comparison rather than selecting participants according to behaviors occurring after assignment.
Suppose 100 students are assigned to an intervention and only 70 engage with it as intended. An intention-to-treat analysis retains the original intervention assignment for all 100 students when estimating the assignment effect.
Students who did not engage are not automatically reclassified as comparison participants.
However, intention-to-treat analysis requires appropriate handling of outcome information. Retaining original assignment labels does not resolve missing outcomes.
When participants who discontinue treatment also fail to complete follow-up assessments, selective loss of outcome data may introduce bias.
Researchers should therefore distinguish nonadherence from missing outcomes, even when they occur together.
Why Can Per-Protocol Analysis Introduce Bias?
A per-protocol effect concerns outcomes under specified adherence to the treatment strategies being compared. Estimating this effect requires more than excluding participants who failed to follow the protocol.
A naive per-protocol analysis may compare only participants who completed their assigned interventions.
However, adherence is not randomly assigned merely because treatment assignment was randomized.
Participants who adhere may differ from nonadherent participants in motivation, health status, prior achievement, resources, or other characteristics influencing outcomes.
For example, students who complete every online learning activity may already be more motivated than those who discontinue participation.
Comparing these selected students with comparison participants can confound adherence-related characteristics with the intervention effect.
Appropriate per-protocol effect estimation may require methods that address time-varying adherence and confounding, such as inverse probability weighting or other causal approaches under defensible assumptions.
Simply removing nonadherent participants does not recreate a randomized experiment among those who remain.
How Does As-Treated Analysis Differ From Per-Protocol Analysis?
An as-treated analysis commonly groups participants according to the treatment they actually received rather than their original assignment.
For example, students assigned to the platform who never use it might be classified as nonusers, while comparison students who obtain platform access might be classified as users.
This approach may appear to reflect actual exposure more closely, but it can sacrifice the protection of randomization.
Actual treatment receipt may depend on motivation, perceived benefit, access, health, or other prognostic factors.
Consequently, an unadjusted as-treated comparison is generally observational, even when the original study used random assignment.
| Analytical approach |
Primary comparison |
Important consideration |
| Intention-to-treat |
Original randomized assignment |
Supports an assignment-effect interpretation under appropriate assumptions. |
| Naive per-protocol |
Participants retained according to observed protocol adherence |
May introduce postrandomization selection bias. |
| As-treated |
Treatment actually received |
May be confounded by factors influencing treatment uptake or switching. |
| Causally adjusted per-protocol estimation |
Specified adherence strategies |
Requires suitable data, methods, and assumptions about adherence-related confounding. |
Analytical labels alone do not establish validity. Researchers should describe the actual comparison, the estimand, and the assumptions required for causal interpretation.
Can Nonadherence Make an Effective Intervention Appear Ineffective?
Yes, under some circumstances.
If an intervention is beneficial when received but many assigned participants do not receive it, the average effect of assignment may be smaller than the effect under sustained receipt.
However, this should not automatically be described as bias. If the intended estimand is the effect of offering the intervention under realistic conditions, low uptake may be part of the effect being estimated.
For example, a digital learning program that produces strong outcomes among highly engaged students but is rarely used may have a modest assignment effect in ordinary institutional practice.
That finding may be important for educational decision-making.
The distinction between intervention efficacy under specified adherence and effectiveness under actual implementation conditions should be defined before interpreting the results.
Can Nonadherence Exaggerate an Apparent Intervention Effect?
Yes. Selective adherence can make an intervention appear more beneficial when researchers analyze only participants who comply.
Suppose academically motivated students are more likely to complete all intervention activities and also more likely to achieve high examination scores regardless of the intervention.
A comparison restricted to completers may attribute their favorable outcomes to the intervention rather than to their preexisting characteristics.
The direction of bias is not guaranteed. Participants may also adhere because they are struggling, experience adverse outcomes, or have other characteristics associated with poorer prognosis.
Researchers should investigate the causal determinants of adherence rather than assume that nonadherence always weakens estimated effects.
How Is Nonadherence Different From Intervention Fidelity?
Intervention fidelity concerns whether an intervention is delivered as intended, including relevant components, dose, quality, and differentiation from comparison conditions.
Adherence commonly concerns whether participants follow or receive the assigned intervention strategy.
For example, a teacher may omit essential instructional components despite students attending every session. This is primarily a delivery fidelity problem.
Alternatively, the teacher may deliver the intervention exactly as planned while some students fail to participate. This is primarily a participant adherence problem.
Both can affect the treatment contrast, but their mechanisms differ.
Researchers should document intervention delivery and participant engagement separately when these distinctions matter to the causal question.
How Should Researchers Measure Intervention Adherence?
Adherence measurement should reflect the intervention's substantive components rather than relying on a convenient but weak proxy.
For a digital learning program, possible indicators include completed learning activities, frequency of engagement, exposure duration, and completion of required instructional sequences.
Platform logins alone may not establish meaningful participation. A student can log in without completing the intended learning activities.
Similarly, attendance does not necessarily establish that a participant received the intended dose or quality of an intervention.
Researchers should define adherence criteria prospectively, explain their rationale, and distinguish continuous measures of exposure from binary adherence classifications.
Arbitrary thresholds created after examining outcomes may introduce analytical flexibility and misleading interpretations.
Can Statistical Methods Estimate Effects Under Full Adherence?
Potentially, but appropriate methods depend on the design and causal estimand.
Methods such as inverse probability weighting, g-computation, and instrumental-variable approaches may help estimate certain adherence-related effects under specific assumptions.
For example, instrumental-variable methods may use randomized assignment as an instrument for treatment receipt when assumptions such as relevance, exclusion restrictions, and suitable monotonicity conditions are defensible for the estimand.
These assumptions are substantive, not merely statistical. They may be difficult to justify when assignment influences outcomes through mechanisms other than the measured treatment receipt.
Similarly, inverse probability weighting requires adequate measurement of relevant confounders, suitable models, and sufficient positivity.
Advanced methods do not automatically recover the effect that would have occurred under perfect adherence.
How Can Researchers Reduce Nonadherence Before It Occurs?
Researchers should first investigate whether the intervention is feasible and acceptable for the intended participants.
Nonadherence may arise from excessive burden, limited access, unclear instructions, technical difficulties, adverse experiences, or conflicts with participants' responsibilities.
Pilot testing can help identify practical barriers before a full evaluation.
Other safeguards may include accessible delivery procedures, appropriate training, clear intervention instructions, technical support, and prospective monitoring of implementation.
Retention and adherence procedures should respect participant autonomy and applicable ethical requirements.
Importantly, researchers should not redefine the intervention after observing poor adherence without acknowledging that the treatment strategy has changed.
What Should Be Reported When Nonadherence Occurs?
Researchers should describe the intended intervention, actual treatment receipt, relevant deviations from the protocol, and the methods used to measure adherence.
They should report the primary estimand and explain how the analysis handles nonadherence.
The CONSORT 2025 reporting framework emphasizes transparent reporting of intervention delivery, adherence, and analytical methods in randomized trials.
When additional per-protocol or as-treated analyses are presented, researchers should distinguish them from the randomized assignment comparison and explain their assumptions and limitations.
Transparent reporting helps readers determine whether the findings concern the effect of offering an intervention, receiving it, or following a specified treatment strategy.
