Handling Treatment Changes After Assignment

Why the Protocol, Not the Drug Label, Determines What Counts as a Change

After treatment assignment, patients may stop the assigned treatment, switch to another treatment, change dose, or add a second treatment. These events are clinically important, but none has a universal statistical classification. A switch is a protocol deviation only if the protocol rules out that switch. Discontinuation is nonadherence only if continuing treatment is required. Adding another treatment may be allowed, prohibited, or explicitly required.

The key is to define the treatment strategy over follow-up before classifying observed histories. An initiation-based strategy specifies whether and when a treatment is started, while leaving subsequent treatment changes unrestricted or governed by a stated usual-care policy. A sustained strategy specifies a treatment course over time, so stopping or switching may violate the strategy. The same observed history can therefore conform to one strategy and depart from another.

Write the Strategy as a Rule for a Treatment History

A useful protocol description answers four concrete questions: What treatment must be initiated, and by when? What must happen after initiation? Are dose adjustments permitted? Are switches, discontinuation, or add-on treatments allowed? “Receive treatment X” is not sufficiently precise if the protocol does not say whether that means start X once, remain on X, or use X with specified co-treatments.

For an initiation-based strategy, the protocol may require starting treatment X within a grace period and then allow later care to proceed as it otherwise would. In that case, later discontinuation, switching, and augmentation do not retroactively undo the fact that the initiation rule was met. This is not a claim that those later events are irrelevant to outcomes. It means that the strategy being compared does not prescribe them.

A sustained strategy instead describes an allowable course. For example, it might require starting X within 30 days, continuing X thereafter, and permitting dose titration and a specified add-on treatment. A switch to Y or discontinuation of X would then be a deviation; a permitted titration or add-on would not. The exact details determine the answer.

Representing Adherence and the First Deviation

For a discrete sequence of decision times, define \(I_t^g=1\) when the action observed at time \(t\) is permitted by strategy \(g\), and \(I_t^g=0\) otherwise. The indicator that the history has remained compatible with the strategy through time \(t\) is

Because each factor is either zero or one, \(C_t^g=1\) exactly when every action from time 0 through time \(t\) was permitted. Once a disallowed action occurs, \(C_t^g\) becomes zero and remains zero. The first deviation time is the earliest \(t\) for which \(I_t^g=0\). With irregularly timed records, the protocol should specify how event dates map to decision times; the recorded visit date need not be the actual date of a change.

This representation is a classification rule, not a causal adjustment method. In an emulation that uses artificial censoring at deviation, the rule identifies when censoring occurs. It does not by itself make those who remain adherent comparable to those who do not.

Proof. If every action through \(t\) is permitted, then \(I_k^g=1\) for all \(k\leq t\), so their product is 1. Conversely, if at least one action through \(t\) is disallowed, then at least one factor is zero, so the product is zero. By definition, there are no disallowed actions before \(t^\ast\), and there is a disallowed action at \(t^\ast\). Thus the product is 1 before \(t^\ast\) and zero at and after \(t^\ast\). \(\square\)

Switching, Discontinuation, and Augmentation in Practice

Treatment records often contain more detail than a single category such as “on treatment.” A patient can stop X but continue a related treatment, take X intermittently, or add Y while remaining on X. Protocols should state whether treatment means a specific drug, a drug class, a dose range, or a regimen consisting of multiple components.

What the Strategy-Specific Outcome Means

Let \(Y^g\) denote the outcome that would be observed under strategy \(g\). Its meaning depends on the complete strategy, including what is allowed after initiation. Under a sustained strategy, \(Y^g\) refers to the outcome under the specified treatment course. Under an initiation-based strategy that leaves later care to usual practice, \(Y^g\) refers to the outcome under the initiation rule together with that downstream-care policy.

A consistency argument links observed outcomes to these potential outcomes: when a participant’s observed treatment history is generated by the strategy being considered, the observed outcome equals that participant’s potential outcome under the strategy. This link requires that the strategy and relevant treatment versions be well-defined. If “usual care” can mean substantially different downstream policies, the intended policy or target population must be clear enough for \(Y^g\) to have a stable interpretation.

Proof. Consistency states that a participant’s observed outcome equals the potential outcome under the treatment history and care policy the participant actually experienced. By assumption, that experienced history and care policy are those specified by \(g\). Substituting the strategy-generated history for the observed history gives \(Y=Y^g\). This equality is an individual-level consistency statement; it does not imply that outcomes among observed adherers identify the population mean \(E[Y^g]\). \(\square\)

For an initiation-based strategy, later switching or discontinuation can be compatible with \(g\), so consistency does not require the participant to remain on the initiated treatment. For a sustained strategy, those same events may make the observed history incompatible with \(g\). In observational data, interpreting a contrast of strategy-specific outcomes also requires appropriate control of confounding at assignment and, when adherence over time matters, of time-varying factors that influence subsequent treatment and outcome. Positivity and a well-defined intervention are also needed. Those assumptions cannot be supplied merely by labeling a history “adherent.”

Diagnostics, Alternatives, and Common Mistakes

A practical protocol audit should compare each treatment-change rule with the recorded data. Check whether initiation dates, stops, switches, doses, and add-on treatments are observed at adequate resolution; whether permitted exceptions have explicit criteria; and whether apparent deviations arise from genuine changes or from missing records. Summarize the frequency and timing of each type of change under the strategy definition. If nearly everyone deviates at a particular time, the sustained strategy may be infeasible in the target population or poorly matched to available care.

There are two common estimand choices, not two interchangeable ways to code the same treatment history. An initiation-based, treatment-policy approach asks about the consequences of adopting an initiation rule while allowing later care to unfold as specified. A per-protocol approach under sustained strategies asks about outcomes if participants followed the specified regimen. The latter typically requires handling deviations, for example by artificial censoring and appropriate adjustment for informative adherence. Simply comparing those who happened to remain on treatment with those who did not can be biased, because health changes may affect both continuation and outcome.

• Mistake: treating every stop as nonadherence. Correction: a stop is a deviation only when the strategy requires continuing treatment.
• Mistake: defining groups by eventual treatment history. For example, calling participants “initiators” only if they start by day 30 uses future information to classify them. Define strategies at time zero and apply the protocol’s grace-period rules instead.
• Mistake: assuming any active drug satisfies a sustained strategy. Correction: specify whether the requirement concerns a particular drug, a class, or a regimen.
• Mistake: treating a permitted add-on as automatically nonadherence. Correction: check whether the protocol allows it, and under what conditions.

The protocol’s precision is a limitation as well as a strength. Very strict sustained regimens may be difficult to emulate when treatment is measured coarsely or exceptions are common. Broad initiation strategies can be more feasible, but they answer a different question and may combine many downstream treatment patterns. No classification rule removes confounding, measurement error, or ambiguity in the intervention itself.