Choosing an Intercurrent-Event Strategy

Start With the Clinical Question

An intercurrent event occurs after treatment assignment and affects either the interpretation or the existence of an outcome. Examples include starting rescue medication, discontinuing assigned treatment, switching therapy, or dying before a planned assessment. The event is not automatically a missing value: an outcome recorded after rescue medication may be observed, but its meaning depends on the question being asked.

The strategy for an event specifies how the clinical question handles outcomes in the presence of that event. It is part of defining the estimand, not a statistical technique for filling in missing values. As established in the earlier material on estimands and intercurrent events, an estimand also specifies the population, treatment conditions, outcome variable, and summary measure. Those attributes must remain clear while deciding how to handle each event.

Definition: An intercurrent-event strategy states how an estimand’s clinical question treats outcomes affected by a specified post-assignment event. The five standard strategies are treatment-policy, hypothetical, composite, while-on-treatment, and principal-stratum. A strategy answers a question; it does not, by itself, prescribe an estimator or solve missing-data problems.

A useful first question is: What should the treatment effect mean if this event occurs? If the answer is “the effect of assignment in the care pathway people actually follow,” treatment-policy may fit. If the question concerns a world where the event did not occur, consider hypothetical. If the event itself is part of success or failure, consider composite. If only outcomes up to the event are relevant, consider while-on-treatment. If the target is a group defined by how people would experience the event under either treatment, consider principal-stratum.

Five Strategies, Five Different Questions

Strategy	Question it represents	Key implication
Treatment-policy	What is the effect of assignment regardless of whether the event occurs?	Outcomes after the event remain relevant to the assigned-treatment comparison.
Hypothetical	What would the effect be if the event had not occurred?	The target is a specified counterfactual scenario.
Composite	What is the effect on an outcome that incorporates the event?	The event becomes part of the outcome definition.
While-on-treatment	What is the effect on outcomes while participants remain on assigned treatment, up to the event?	Post-event outcomes do not contribute to the defined measure.
Principal-stratum	What is the effect in a group defined by its event status under the relevant treatment conditions?	Group membership can depend on unobserved counterfactual event status.

These are not five competing ways to estimate one universally defined effect. They may define different effects because they answer different questions. The choice should follow clinical relevance, the intended decision, and whether the outcome under the strategy can be meaningfully defined. Convenience, the frequency of an event, or the availability of a particular analysis should not determine the strategy on its own.

For compact notation in the examples, let \(A\) denote randomized assignment, with \(A=1\) for the new treatment and \(A=0\) for control. Let \(Y(a)\) denote the outcome under assignment \(a\), and let \(E(a)\) indicate whether the event would occur under that assignment. These are potential outcomes: only one assignment is observed for any one participant. The examples use a mean difference as the summary measure unless otherwise stated.

Worked Example: Rescue Medication and the Treatment-Policy Question

Question. A hypothetical randomized trial compares two treatments for pain. Some participants take rescue medication before the week-8 pain assessment. The clinical team wants to know the effect of assigning each treatment in a care pathway where participants may use rescue medication as needed.

This question includes the consequences of rescue use rather than asking what pain would have been without rescue. The treatment-policy strategy therefore fits. Suppose the week-8 mean pain score, on a scale where lower is better, is 4.8 for the 100 participants assigned the new treatment and 5.5 for the 100 assigned control. Assume the week-8 outcome is defined and observed for everyone, including after rescue use.

The assigned-treatment mean difference is \( \bar{x}_1-\bar{x}_0=4.8-5.5=-0.7 \) points.

Thus, in this hypothetical example, assignment to the new treatment is associated with a mean week-8 pain score 0.7 points lower than assignment to control under the rescue-medication policy used in the trial. The result includes outcomes following rescue use; it is not an estimate of the effect that would have occurred if nobody had used rescue medication. Randomization supports the assigned-group comparison, while the stated availability and measurement of post-rescue outcomes make the specified outcome usable here.

Worked Example: Discontinuation and a Hypothetical Question

Question. In a hypothetical trial, some participants stop assigned treatment because of adverse effects. Investigators ask what the mean week-12 outcome would have been if participants had continued the assigned treatment without stopping. This is a hypothetical question: it specifies a counterfactual scenario in which discontinuation does not occur.

Suppose the target is the week-12 outcome under that no-discontinuation scenario, and an analysis under stated modeling assumptions estimates means of 12.0 units under the new treatment and 10.5 units under control. The hypothetical contrast is \(12.0-10.5=1.5\) units.

The example’s estimated 1.5-unit difference describes the specified no-discontinuation scenario, not necessarily the effect of assignment under the trial’s actual discontinuation experience. Its interpretation depends on whether the scenario is clinically meaningful and whether the counterfactual outcomes can be identified from the available data and assumptions. For instance, if stopping is related to worsening outcomes, simply averaging observed outcomes among participants who continued would not automatically recover the hypothetical means. The numerical contrast illustrates the target; it does not establish that the assumptions needed to estimate it are true.

Worked Example: Making an Event Part of the Outcome

Question. In a hypothetical trial for a serious condition, clinicians regard either being alive without hospitalization at day 30 or having been hospitalized by day 30 as meaningfully different outcomes. They decide that hospitalization should count as failure, rather than asking only about a symptom score after hospitalization.

Define a binary composite outcome \(Z\) as success only if the participant is alive and has not been hospitalized by day 30. Suppose 62 of 100 participants assigned the new treatment and 50 of 100 assigned control meet this definition. The estimated success proportions are \(62/100=0.62\) and \(50/100=0.50\). The risk difference is \(0.62-0.50=0.12\), or 12 percentage points.

Under this composite definition, the new-treatment group has a 12-percentage-point higher observed success proportion in this hypothetical data. The event is included in the outcome itself. The result is not a 12-point improvement in a symptom score, nor does it say that every component improved. A composite is appropriate only if the event and other outcome components can be combined in a clinically interpretable way; the definition and component results should be transparent.

Worked Example: Choosing Between While-on-Treatment and Principal-Stratum Questions

Question. A hypothetical trial records a symptom score during assigned treatment, but some participants discontinue before week 8. If clinicians want to describe symptom burden only during exposure to assigned treatment, a while-on-treatment strategy may fit. Suppose the prespecified summary is the mean of scheduled symptom scores obtained before discontinuation, averaged over eligible participant-visits. In the illustrative data, the mean is 3.1 points under the new treatment and 3.8 under control. The contrast is \(3.1-3.8=-0.7\) points, with lower scores indicating fewer symptoms.

This summary concerns scores during assigned treatment as defined; it does not describe a common week-8 outcome for everyone. In particular, if time on treatment differs between groups, the sets of contributing visits may differ. The measure and its weighting must therefore be clinically interpretable, and the observed pre-discontinuation data must be adequate for the intended summary.

A different question would ask for the treatment effect among people who would avoid discontinuation under either assignment. That is a principal-stratum question: the target group is defined by \(E(1)=0\) and \(E(0)=0\), where \(E(a)=1\) denotes discontinuation under assignment \(a\). Suppose, purely for illustration, a model with explicit assumptions estimates a mean difference of \(-0.5\) points in this group. That number targets the always-nondiscontinuing stratum, not all randomized participants and not simply those observed to continue. For each participant, only one of \(E(1)\) and \(E(0)\) is observed, so membership in this stratum is generally not directly known. The hypothetical estimate must not be treated as identified merely because the stratum can be named.

Why the Choice Changes the Result

The distinction can be seen by writing the target for each strategy. For a treatment-policy question with a mean outcome, the contrast is based on outcomes under assignment, including the event’s consequences:

\Delta_{\mathrm{TP}}=E\{Y(1)\}-E\{Y(0)\}.

For a hypothetical question, the outcomes refer instead to a specified scenario in which the event is prevented. Write \(Y(a;\text{no }E)\) for the outcome under assignment \(a\) in that scenario. Then the target is

\Delta_{\mathrm{H}}=E\{Y(1;\text{no }E)\}-E\{Y(0;\text{no }E)\}.

For a composite strategy, define a new outcome \(Z(a)\) that incorporates the event and the clinical outcome; the contrast is \(E\{Z(1)\}-E\{Z(0)\}\). For a principal-stratum question, the mean contrast is conditional on a stratum defined by potential event status, such as

\Delta_{\mathrm{PS}}=E\{Y(1)-Y(0)\mid E(1)=0,\ E(0)=0\}.

These expressions show why strategies cannot be swapped without changing the question. The first averages over the event’s actual consequences under each assignment. The hypothetical target changes the world being described. The composite changes the outcome variable. The principal-stratum target changes the population. A while-on-treatment strategy likewise defines the outcome using measurements before the event, so the measure and its summary must say exactly which observations contribute.

A basic justification for keeping these definitions separate is that randomization balances assigned groups, but it does not make post-randomization event groups exchangeable. For example, participants who discontinue may differ in prognosis from those who continue. Comparing observed continuers across treatment groups can therefore compare selected groups rather than estimate the effect in everyone or in a well-defined principal stratum. Randomization alone does not reveal both potential event statuses for any participant.

A Practical Decision Process

Name the event and outcome.
Specify when the event occurs and which outcome is affected. Do not use “dropout” as a substitute for a precisely defined event or assume an unrecorded outcome is itself an event.

State the decision question.
Ask whether the goal concerns actual care after the event, a world without it, a combined success/failure outcome, outcomes only before it, or a group defined by potential event status.

Check that the target is interpretable.
Define the outcome, time point or period, population, and summary. For a hypothetical strategy, specify the no-event scenario; for a composite, define its components; for a principal stratum, state its membership rule.

Ask what must be assumed to estimate it.
Consider whether outcomes are measured after the event, whether outcomes under a hypothetical scenario are supported by data, whether pre-event observations are sufficient, or whether principal-stratum membership is identifiable under defensible assumptions.

For example, if the clinical decision is whether to prescribe a treatment knowing rescue medication will remain available, a treatment-policy question may be directly relevant. If the decision concerns biological efficacy absent rescue, a hypothetical question may be closer to the goal, but it requires a credible no-rescue scenario. Neither framing is automatically superior; the decision context determines which is useful.

Assumptions, Diagnostics, and Failure Modes

The strategy defines the target, while the estimation method supplies additional assumptions. A treatment-policy target still requires outcome data or an appropriate missing-data analysis when post-event outcomes are unobserved. The strategy does not make missing outcomes disappear. A hypothetical target needs assumptions linking observed information to outcomes under the specified event-free scenario; those assumptions should be explicit and examined through sensitivity analyses when plausible alternatives matter. A principal-stratum target requires particular care because stratum membership depends on counterfactual event status.

Useful checks include whether event definitions and timing are consistent across treatment groups; whether post-event outcomes were collected when the strategy needs them; whether the composite’s components have compatible clinical meaning; and whether a while-on-treatment summary gives comparable, meaningful weight to the observation periods. For hypothetical or principal-stratum analyses, assess whether estimates are highly dependent on untestable assumptions and whether conclusions change under reasonable alternatives.

Strategy mistaken for missing-data handling. Calling an analysis treatment-policy does not justify ignoring unobserved outcomes after rescue or discontinuation.
Hypothetical interpreted as observed. Outcomes in a no-event scenario are counterfactual when the event occurred; they are not directly recovered by deleting affected participants.
Composite treated as one ordinary component. A favorable overall composite does not imply that each component improved. State the composite definition and inspect its components.
While-on-treatment interpreted as an all-participant fixed-time effect. Measurements before discontinuation may cover different periods. Define the summary and do not quietly present it as a common-time outcome.
Observed subgroup called a principal stratum. Those observed not to discontinue under their assigned treatment are not necessarily the people who would avoid discontinuation under both treatments.

A common mistake is to choose whichever strategy yields the most favorable result, then describe it as “the treatment effect.” The correction is to define the clinical question before examining the comparative result, specify the strategy and its assumptions, and interpret the estimate only for that target. Different strategies may give different estimates without contradiction because they can represent different outcomes, scenarios, or populations.

Key takeaway: Choose an intercurrent-event strategy by asking what the clinical decision needs the treatment effect to mean. Then specify the target precisely and check whether the data and assumptions support its estimation. No strategy is universally best, and none automatically resolves missing data.

Check Your Understanding

For each item, identify the best-fitting strategy for the stated question and compare your reasoning with the solutions below.

A trial asks for the effect of assignment when participants may take rescue medication, and it collects the outcome after rescue. Which strategy fits?
Investigators ask what the outcome would have been if treatment discontinuation had not occurred. Which strategy fits, and what makes this a demanding question?
A study defines success as being alive and free of hospitalization. Which strategy handles hospitalization, and what must be made clear when interpreting the result?
Why is a comparison of participants observed to continue treatment not automatically a principal-stratum effect?
A team reports a symptom mean using only measurements collected before assigned treatment stopped. Name the likely strategy and one interpretive check.

Solutions. (1) Treatment-policy: the question includes the consequences of rescue use, and post-rescue outcomes remain part of the outcome. (2) Hypothetical: it targets a specified no-discontinuation scenario, whose outcomes are counterfactual for participants who discontinued and require defensible identification assumptions. (3) Composite: hospitalization is incorporated into the success definition; state the components and do not infer that every component improved from the overall result alone. (4) Continuation is observed under only the assigned treatment, whereas a principal stratum can be defined by event status under both assignments; observed continuers need not be the always-continuing group. (5) While-on-treatment; check that the observation window and summary are meaningful, especially if time on treatment differs between groups.