Representing Nonadherence and Treatment Discontinuation

Why Dose-Taking Behavior Belongs in a Trial Simulation

A prescribed regimen is not necessarily the regimen a participant receives. A participant may miss an occasional dose, pause treatment for several days, reduce doses informally, or permanently discontinue the assigned drug. These behaviors change the input to the pharmacokinetic and pharmacodynamic models, and therefore may change the outcomes a trial is designed to measure. A simulation that gives every participant every scheduled dose can misrepresent both exposure and the performance of a trial design when departures from the regimen are plausible.

The simulation should represent what happens to dosing separately from whether an outcome is observed. Discontinuation is not itself dropout: participants may stop study drug but continue scheduled assessments, and their later outcomes remain relevant to the treatment strategy they were assigned. The earlier material on dropout and informative missingness addresses missing outcome records; here the focus is on dose histories and their consequences.

An interruption can be represented by a run of zero dose indicators, with a separate process determining when the run begins and ends. Discontinuation can be represented by a stopping time \(T_i^{\mathrm{disc}}\): after that time, assigned-drug doses are zero. If the protocol permits rescue medication, background therapy, or a restart, those actions require their own rules. A safety hold or restart should follow the relevant protocol rule rather than being silently treated as ordinary nonadherence.

Specify the Behavior Process Before Simulating It

The simplest model makes dose-taking decisions independently with a constant probability \(q\): \(A_{ik}\sim\mathrm{Bernoulli}(q)\). This may be a useful starting scenario, but it assumes that each dose decision is independent of prior decisions and that the probability of taking a dose is the same for every participant and occasion. Those assumptions are often too strong. Adherence may depend on treatment burden, participant characteristics, adverse events, perceived benefit, prior missed doses, or access to medication.

A more realistic specification models adherence conditionally on information available at that time. For example, a probability model may use participant covariates and recent history to determine the chance of taking the next dose. An alternative is a transition model: the probability of a missed dose may be higher after a missed dose, allowing missed doses to cluster. Temporary interruptions can be simulated with start and restart rules, while permanent discontinuation can have a separate event model. The key is temporal order: information generated in the simulation should influence only subsequent behavior, not earlier dose decisions.

The event model should also distinguish causes when they imply different future histories. A safety-related hold may end after a stated recovery condition; a participant who stops because of perceived lack of benefit may be less likely to restart. If the simulation includes these mechanisms, it must state how they operate and how they affect both dosing and follow-up. A model that makes adherence independent of response and safety is a defensible scenario only when that independence is intended, not a default truth.

To represent dependence, one can instead define probabilities conditional on the preceding state. For instance, let the next dose be missed with probability 0.10 after a taken dose and 0.40 after a missed dose. The resulting process allows missed doses to cluster. A simulation should be explicit about whether it uses independent dose decisions, a history-dependent model, an interruption process, or a combination.

Translate the Dose History Into Exposure

The dose-taking process matters only through its consequences for treatment and outcomes. For each simulated participant, pass the administered dose history—not merely the prescribed regimen—into the PK model. Then use the resulting exposure history in the PD or exposure-response model. A missed dose does not necessarily cause an immediate proportional reduction in concentration: the magnitude and timing of the change depend on elimination, absorption, dosing interval, and the model’s state at the time of the omission.

A short justification follows directly from the model’s state update. At each occasion, the previous concentration contributes the fraction \(rC_k\) that remains, and the new administered dose contributes \(1.00A_{k+1}\). When \(A_{k+1}=0\), the new-dose term is zero but the existing concentration does not disappear. Repeating that update through an interruption gives the concentration trajectory; restarting adds doses to the residual concentration left at restart.

Discontinuation, Follow-Up, and the Outcome Strategy

After permanent discontinuation, assigned-drug dosing stops, but the participant’s clinical course need not stop. Drug concentrations may decline gradually; symptoms or disease status may persist or change; rescue treatment may begin; and scheduled measurements may continue. A trial simulation should therefore specify post-discontinuation PK/PD and care as appropriate to the intended outcome. Setting all outcomes after discontinuation to missing confuses an exposure event with a data-collection event.

Two outcome strategies are particularly important to distinguish. Under a treatment-policy strategy, the outcome is evaluated regardless of missed doses or discontinuation, including the consequences of subsequent care as specified by the trial and model. Under a hypothetical sustained-exposure strategy, the outcome is defined under a hypothetical intervention in which the assigned treatment is taken as specified throughout the relevant period. The second is not the observed outcome among only those who adhered: selecting adherent participants can select a different mix of prognosis, response, and tolerability.

Simulate Discontinuation as an Event Process

A discrete-time discontinuation model can assign a probability of stopping in each interval among participants still on treatment. Let \(h_k\) be the conditional probability of discontinuation during interval \(k\), given that the participant has not previously discontinued. If the hazards are fixed for this example, the probability of remaining on treatment through \(K\) intervals is the product of the conditional probabilities of not stopping:

This product follows by applying the definition of conditional probability successively: survival through interval \(k\), conditional on survival through \(k-1\), has probability \(1-h_k\). Multiplying these conditional probabilities gives the probability of surviving all \(K\) intervals. If hazards depend on evolving safety, response, or adherence history, use the corresponding conditional \(h_{ik}\) for each participant and interval rather than a fixed common value.

Diagnostics, Alternatives, and Limitations

Diagnostics should examine the behavior process as well as its downstream implications. Review the proportions of doses taken, missed, and interrupted; the distribution of run lengths for missed doses; time to discontinuation; and these summaries by relevant participant characteristics and simulated histories. Check whether simulated probabilities are in range and whether events occur in the intended order. Then inspect resulting exposure trajectories, including accumulation, decline during interruption, and washout after stopping. Implausible exposure patterns may indicate an implementation error or an inconsistent link between dose records and the PK model.

A constant independent probability is transparent and easy to stress-test, but it cannot represent clustering or behavior linked to response and tolerability. A history-dependent model can represent clustering, but may require more assumptions and data to calibrate. A rule-based protocol model is often appropriate for mandated safety holds and restarts; it should not be used to imply that all participant-driven nonadherence follows a deterministic rule. More detailed event-time models can represent variable timing, but their additional complexity is useful only if the relevant parameters and consequences are supportable.

The principal limitation is that adherence is difficult to predict exactly, particularly when it depends on unmeasured or changing motivations. A single “best guess” process can therefore create false precision. Use plausible alternative adherence and discontinuation scenarios where uncertainty could affect the design decision. Also distinguish the model’s treatment-policy predictions from predictions under sustained exposure: the latter require counterfactual assumptions and may be sensitive to how response, safety, and post-discontinuation care are represented.

• Pitfall: Treating discontinuation as dropout and deleting every later outcome. Correction: Stop assigned-drug dosing at discontinuation, but continue outcome simulation and observation when the trial collects follow-up data.
• Pitfall: Multiplying the prescribed dose by an average adherence percentage and using that reduced dose at every occasion. Correction: Simulate dose occasions or interruptions when timing matters; the same average number of doses can produce different exposure trajectories.
• Pitfall: Calling outcomes among simulated adherers “the sustained-exposure effect.” Correction: Define the hypothetical sustained regimen and simulate outcomes under that intervention; adherers may differ systematically from nonadherers.