What to Do When You Detect Non-Adherence in a Clinical Trial

Responding to non-adherence in a clinical trial means detecting a dosing deviation while the participant is still in the study, deciding whether it crosses a threshold that warrants contact, establishing why it happened, and documenting the action taken and the dosing that followed. Almost everything written about adherence covers how to measure it. Very little covers what a study team should actually do once a deviation is in front of them, and that gap is where most of the avoidable cost sits. This guide covers the response: the thresholds, the contact, the bias questions a protocol reviewer will raise, and the record you should have at the end.

It assumes you have a monitoring method capable of surfacing a deviation while there is still time to act. If your monitoring is retrospective, pill counts or diaries reviewed at the next visit, the response options in this guide are mostly closed to you, and our guide on how to monitor medication adherence in a clinical trial covers the trade-offs between methods.

Why the response window decides the cost

The cost of a non-adherent participant is not fixed at the moment they miss a dose. It is fixed at the moment you can no longer do anything about it. A participant whose dosing slips in week three and is contacted in week three can usually be recovered: the cause is identified, the regimen resumes, and the deviation is a documented event inside an otherwise evaluable dataset. The same participant discovered at a visit six weeks later, or at database lock, may be a non-evaluable subject, a protocol deviation for the file, and a recruitment cost to repeat.

ICH E9(R1) made this concrete by requiring sponsors to define estimands and to pre-specify how intercurrent events, which include non-adherence and treatment discontinuation, will be handled in the analysis. Every strategy available under E9(R1) is a way of coping with non-adherence after it has happened. A hypothetical strategy asks what the effect would have been had the participant adhered, and pays for the question in assumptions. A per-protocol style approach excludes the participant and pays in power and in bias. A treatment policy strategy keeps the participant in and pays in a diluted effect estimate. The statistics are well developed, and none of them return the information that was lost. The only strategy that does is preventing the intercurrent event from completing, which is a monitoring and response question, not an analysis question.

The same logic now appears on the quality side. ICH E6(R3) expects sponsors to identify the issues that matter to the reliability of trial results, to detect them as they emerge, and to respond and record the response. Dosing that has drifted from the protocol is close to a canonical example of an issue that matters to result reliability. A trial that can show it detected dosing issues, acted, and recorded the outcome has an answer at inspection that a stack of returned-bottle counts does not provide.

What counts as a deviation worth acting on

Not every irregular dispense deserves a phone call, and a response plan that treats every late dose as an incident will exhaust the site and annoy the participant. The taxonomy that most of the adherence literature now uses (Vrijens et al., 2012) separates non-adherence into three phases: initiation, whether the participant ever starts dosing; implementation, how well actual dosing matches the prescribed regimen while they are taking it; and persistence, whether and when they stop. Each phase has a different response, and a threshold plan should address them separately.

Failures of initiation are the cheapest to fix and the most embarrassing to miss. A participant who is randomised and never doses is fully recoverable in the first days and unrecoverable shortly after, so the threshold here is simple: no first dispense within the expected window after dispensing visit is a contact, without exception.

Implementation deviations need more judgement. A single dose taken a few hours late is noise in most regimens and acting on it teaches the participant that the monitoring is oppressive. A missed dose followed by normal dosing is usually noise too. What is rarely noise is a run: consecutive missed doses, a weekend pattern, or a widening drift in dose timing. The literature calls a run of three or more consecutive omitted days a drug holiday, and drug holidays are worth treating as the primary implementation trigger because they are both common and consequential. Where the threshold sits for your study depends on the pharmacology: a drug with a long half-life forgives a missed day that a short half-life drug does not, and a regimen with a narrow dosing window turns timing drift into a real deviation rather than a cosmetic one. The threshold should be set with the clinical team against the drug’s forgiveness, written into the monitoring plan, and applied uniformly.

Persistence failures usually announce themselves as implementation failures first. A participant heading towards discontinuation rarely stops cleanly; dosing becomes intermittent, holidays lengthen, and then it ends. This is the strongest argument for pattern-level triggers rather than single-event triggers: the pattern is the early warning, and retention interventions work better before the participant has mentally left the study.

Establish the cause before correcting the behaviour

The reflex response to a missed-dose alert is a reminder, and the evidence for reminders alone is modest. Reminder systems address forgetting, and forgetting is only one cause among several. Participants miss doses because of side effects they have not reported, because they misunderstood the regimen, because the kit ran out or travel separated them from it, or because something in their life displaced the trial for a week. Each of these has a different fix, only one of which is a reminder, and two of which (unreported adverse events and regimen misunderstanding) are things the study team urgently needs to know for reasons well beyond adherence.

Even where forgetting is the cause, a reminder only works if it arrives at a moment the participant can act on it. The example I use is my own phone alerting me to a dose while I am changing my toddler: the alert is perfectly accurate and completely ignored, and by the time my hands are free it has scrolled out of sight. A reminder schedule fixed at study setup cannot know which moments those are for any given participant, and someone whose alerts repeatedly land at the wrong time learns to swipe them away, which leaves them worse off than no reminder at all. Continuous monitoring changes what is possible here, because the question stops being whether the prompt fired and becomes what the dosing did next: a dose taken late that evening is a day that recovered on its own, while a miss that becomes a run is a participant who needs a person. Support can then stay light and stay out of the way of participants who are managing fine, and concentrate on the ones who are not, without asking anyone to rearrange their life around the monitoring.

This is why the response should be a conversation rather than a nudge. The alert tells the site team that dosing has slipped and when; the call or message establishes why; the fix follows from the why. A short structured questionnaire, sent to the participant when a threshold trips and positioned as information gathering rather than correction, is a useful first step in that sequence. It reaches the participant while the event is days old rather than weeks, which sidesteps the recall problem that makes visit-based self-report unreliable, and it hands the site team a cause alongside the alert rather than leaving them to open the call blind. It also avoids the confrontation dynamic that retrospective methods create: a participant asked at a visit to explain a discrepant pill count has an incentive to defend the record, while a participant asked two days after a missed weekend what got in the way tends to just answer.

Responding without biasing the trial

A protocol reviewer will raise two objections to in-flight adherence intervention, and both deserve straight answers.

The first is behavioural: contacting a participant about their dosing changes their dosing, so the monitoring is no longer passive observation. This is true, and in a course-correction design it is the point rather than a flaw. The honest way to handle it is to pre-specify it: the escalation pathway, the contact thresholds, and the script live in the protocol or monitoring plan, apply identically across arms, and are reported with the study. What creates bias is not intervention but asymmetric or improvised intervention. A trial in which every participant crossing a pre-specified threshold receives the same structured contact has changed the estimand slightly (the treatment effect is now estimated under supported dosing) and has done so transparently. A trial in which site staff informally chase whichever participants they happen to worry about has introduced something no analysis can adjust for.

The second is blinding. Adherence data at the participant level does not unblind treatment allocation, and the response pathway should keep it that way: site teams act on dosing behaviour, not on anything that could correlate with arm. It is worth stating in the plan that alert thresholds and contact scripts are identical across arms so that the intervention cannot become a channel for differential attention.

There are designs where in-flight correction is the wrong choice. A study whose purpose is to observe naturalistic dosing behaviour should monitor silently, and a regulator may take a view on intervention in a pivotal efficacy trial that they would not take in Phase 2. The decision belongs in protocol design, not in an SOP written after first patient in. What does not vary by design is the value of knowing: even a trial that chooses not to intervene is better off detecting the deviation when it happens and holding the record.

The record is the deliverable

Whatever the response, the trial should end holding a record with three parts for every deviation that crossed a threshold: what was detected and when, what the site did, and the dosing that followed. This record earns its keep three times. At inspection, it is direct evidence of the risk-based quality management that E6(R3) asks for. At analysis, it lets statisticians handle intercurrent events with dated facts rather than reconstruction. And commercially, it changes what an adherence programme is worth: a monitoring method that produces only a percentage tells you how much risk you carried, while a detection and response record shows what the risk was and what was done about it, whichever way the trial went.

There is a version of this record for clean trials too. A study in which no threshold ever tripped is not a study with nothing to show; it is a study with continuous, dated evidence that dosing held, which very few trials can currently produce.

Where the monitoring method decides what is possible

Everything above depends on one property of the monitoring method: the deviation has to reach the study team while the participant is still recoverable. Pill counts, diaries, and refill records report at the cadence of site visits, so the response options they support are retrospective by construction. Electronic monitoring closes the gap only if the data leaves the device without waiting for the participant to sync an app or attend a visit. Pill Connect was built around this requirement: each pill dispense is sensor-verified and time-stamped, the data moves in the background with nothing for the participant to do, and when dosing crosses a threshold your site team receives an alert with the detail needed to act, with the action and outcome captured alongside the dosing record. The monitoring devices that surface these deviations, and how much each can actually tell you, are covered in our guide to the smart pill bottle. How that compares with other methods, and how it holds up in decentralised designs, is covered in the monitoring methods guide and the decentralised trials guide.