Choosing Strategies for Intercurrent Events

One Event, Several Different Questions

An intercurrent event occurs after treatment assignment and affects either the interpretation or existence of the outcome relevant to the clinical question. Treatment discontinuation because of toxicity is one example. Suppose a randomized trial compares a new therapy with control for a chronic condition, and the outcome is improvement in symptoms at week 24, measured in points, with larger values indicating greater improvement. Some participants discontinue assigned treatment before week 24. The clinical question might sound like “Does the new therapy improve symptoms?” But that question does not yet specify what to do with outcomes after discontinuation.

A strategy for handling the event is part of the estimand: it determines which outcome or population the treatment contrast concerns. The strategies below can use the same treatment comparison and clinical setting yet target different quantities. They are not alternative calculations of one fixed effect; choosing among them changes the question being answered.

Potential outcomes are indexed by the treatment assignment a participant would receive; \(Y(1)\) and \(Y(0)\) cannot both be observed for the same participant. The definitions also require treatment conditions, the event, outcome, population, and summary measure to be sufficiently precise. For example, “discontinuation” should specify which treatment is discontinued and when; “symptom improvement” should specify its scale and assessment time.

An event strategy is not a missing-data method. If a participant discontinues treatment but week-24 symptoms are still measured, the outcome is observed and the event strategy determines how it enters the estimand. If the week-24 measurement is unavailable, that is missing outcome data; additional assumptions and methods are needed to address it.

Five Strategies and Their Targets

For a treatment-policy strategy, the outcome is considered regardless of whether the intercurrent event occurs. The contrast includes the consequences of discontinuation and subsequent care as they arise under each assigned treatment. It asks about assignment to each treatment condition in the trial context, not about the isolated biological effect of remaining on treatment.

A hypothetical strategy asks what the outcome would have been under a specified scenario in which the event did not occur. “No discontinuation” must describe a coherent hypothetical intervention: for example, remaining on assigned treatment while otherwise receiving the stated care. The resulting contrast concerns those hypothetical outcomes, not the outcomes actually observed after discontinuation.

A composite strategy incorporates the event into the outcome itself. For example, define success as achieving a prespecified symptom response and not discontinuing assigned treatment by week 24. The target is then a contrast in the probability of this joint success, not a contrast in mean symptom improvement alone.

A while-on-treatment strategy uses outcome information while participants are receiving assigned treatment and defines the target in relation to that period. To make the summary explicit, one could average improvement over each participant’s time on assigned treatment, ending at discontinuation or week 24. This is not the same as a week-24 outcome: participants may contribute over different durations, and an outcome summary must specify how those durations are handled.

A principal-stratum strategy targets a subgroup defined by potential event status under both assignments. In this example, the “always non-discontinuers” are participants who would not discontinue under either treatment. Membership depends on the pair \((D(1),D(0))\), although only one of those event statuses is observed per randomized participant.

Worked Examples

Why Conditioning on the Event Changes the Question

The treatment-policy mean can be decomposed by event status within each treatment condition. For assignment \(a\), the law of total expectation gives:

The treatment-policy target includes both event groups, weighted by their proportions under that assignment. By contrast, comparing only participants who did not discontinue in each randomized group compares \(E[Y(1)\mid D(1)=0]\) with \(E[Y(0)\mid D(0)=0]\). These are different groups: discontinuation may be affected by treatment and by prognosis. Conditioning on the post-randomization event can therefore break the balance created by randomization. It does not, by itself, estimate either the treatment-policy effect or the principal-stratum effect.

The principal stratum avoids defining its target by observed event status in one arm, but it creates a different challenge: both potential event statuses are needed to determine membership, and each participant reveals only one. The hypothetical strategy also refers to unobserved outcomes under a scenario that did not occur. Neither strategy becomes identified merely by naming it.

Assumptions, Diagnostics, and Choice

A clear estimand begins with well-defined treatment conditions, population, outcome, event, and summary. Consistency requires that the outcomes represented by each treatment condition match the treatment actually assigned and received as specified. For a randomized treatment-policy comparison with complete outcome ascertainment, randomization supports comparison of the assignment groups in the target population. If outcomes are missing, assumptions about their distribution and missingness are also needed.

Additional assumptions depend on the strategy. A hypothetical estimand requires a clinically coherent no-event scenario and assumptions adequate to learn its outcomes from observed data. A composite estimand requires a prespecified rule for combining the event and clinical outcome. A while-on-treatment estimand requires a defined treatment exposure period and outcome summary; event-related truncation can make the available outcomes differ systematically across participants. A principal-stratum estimand requires a scientifically useful stratum and assumptions or design information sufficient for identification; the stratum is not directly observed.

Diagnostics should first check whether the event and outcome are defined and recorded as intended: inspect event frequencies by randomized group, event timing, post-event outcome availability, and reasons for discontinuation. Large differences in event rates can make the choice of strategy especially consequential. For hypothetical and principal-stratum analyses, also examine whether conclusions depend strongly on assumptions that cannot be checked from observed data alone. A precise numerical result does not resolve a poorly defined or weakly supported target.

In an adaptive trial, changing allocation or making another planned adaptation does not automatically change the estimand. But the target is preserved only if the treatment conditions, population, outcome, event strategy, and summary continue to mean what the estimand specifies. A change that alters who is studied, what treatment is assigned, or what is counted after an event may change the question unless the estimand is explicitly maintained.

Common Mistake

Wrong reasoning: “The participants who did not discontinue are the ones who stayed on treatment, so comparing them gives the randomized treatment effect among people who can tolerate treatment.” The observed event-free groups need not be comparable: treatment assignment can affect discontinuation, and underlying prognosis can affect both discontinuation and outcome. The comparison conditions on a post-randomization variable.

Correction: State the intended target first. Use all randomized participants and outcomes regardless of event for a treatment-policy question; define a no-event scenario for a hypothetical question; incorporate the event explicitly in a composite; specify the time-dependent outcome for a while-on-treatment question; or define the principal stratum using potential event status under both assignments. Do not treat these targets as interchangeable.