Why Multiple Events Need a Joint Specification
A trial participant may discontinue assigned treatment, receive rescue medication, and later begin another therapy before the endpoint is measured. These events do not necessarily raise the same clinical question. Discontinuation may be part of the treatment experience the trial intends to evaluate; rescue may count as failure; and subsequent therapy may prompt the question of what the outcome would have been without it. Treating all three events with one blanket rule can therefore produce an estimand that does not answer the intended question.
An estimand specifies the treatment effect of interest, including how intercurrent events are handled. With multiple events, the specification must say not only what strategy applies to each event, but also what happens when events overlap or occur in sequence. The strategy combination is part of the clinical question, not merely a convenient way to organize an analysis.
The available strategies include treatment policy, hypothetical, composite, principal stratum, and while-on-treatment approaches. Each has a different interpretation. For example, a treatment-policy strategy uses outcomes regardless of the event; a hypothetical strategy targets an outcome under a stated counterfactual condition; and a composite strategy incorporates the event into the endpoint definition. These strategies should not be substituted for one another just because the data are incomplete or difficult to analyze.
Writing the Joint Strategy Precisely
Let \(a\) denote an assigned treatment, and let \(Y^{a,g}\) denote the endpoint under treatment condition \(a\) and a specified joint event-handling strategy \(g\). For a continuous endpoint, one possible treatment contrast is
Here \(g\) includes the strategy for every event in scope. The contrast is interpretable only after the endpoint and event strategies have been made clear. A negative value, for example, means a lower population mean under treatment condition 1 than under treatment condition 0 when lower endpoint values are favorable.
A practical specification should answer four questions. First, which events count, and over what time window? Second, what strategy applies to each event? Third, can more than one event occur for the same participant, and if so, which rule determines the endpoint? Fourth, what outcomes are needed after each event to identify or estimate the target? These questions prevent an event-specific rule from silently overriding another one.
Define each event operationally, including its start date or assessment window, and whether it may occur more than once.
State the clinical meaning of the strategy, such as using outcomes regardless of an event, targeting an outcome under a counterfactual condition, or incorporating the event into the endpoint.
Specify whether strategies act jointly, whether one event takes precedence, or whether an endpoint rule is applied at the first qualifying event.
Confirm that the endpoint is defined for each relevant event pattern and identify the post-event outcome information or assumptions needed to estimate it.
If event histories are represented by \(H\), the overall mean under a joint strategy can be written as a sum over possible histories:
This follows by partitioning the population into mutually exclusive event-history groups and applying the law of total expectation. The identity is a useful bookkeeping device: every relevant history contributes to the overall mean. It does not, by itself, identify the conditional mean for a hypothetical outcome. That may require additional assumptions and a suitable model.
Worked Example: Specify Strategies for Three Events
Worked Example: Specify Strategies for Three Events
Suppose a hypothetical randomized trial compares a new treatment with control in adults with a chronic condition. The endpoint is symptom burden at week 24 on a scale from 0 to 10, where higher scores are worse. Participants may discontinue assigned treatment, receive rescue medication, or initiate subsequent therapy before week 24. The clinical team wants to evaluate symptom burden while allowing treatment discontinuation as it occurs, treating rescue as a poor outcome, and asking what symptom burden would have been without subsequent therapy.
One coherent specification is:
- Discontinuation: treatment-policy strategy. Use the week-24 symptom score whether or not the participant discontinued assigned treatment.
- Rescue medication: composite strategy. If rescue is received before week 24, assign the endpoint value 10, the worst value on the defined scale.
- Subsequent therapy: hypothetical strategy. For participants who initiate subsequent therapy without having received rescue, target their week-24 score under the condition that they did not initiate that subsequent therapy.
The overlap rule is explicit: rescue takes precedence over the hypothetical subsequent-therapy strategy. Thus, for a participant who receives rescue and later initiates subsequent therapy, the endpoint is 10. For a participant who initiates subsequent therapy but has not received rescue, the endpoint is the counterfactual week-24 score without that therapy. Discontinuation alone does not change how the endpoint is defined.
This combination answers a specific question. It does not describe the effect of preventing every event: discontinuation is included as part of the treatment experience, rescue is counted as failure, and subsequent therapy is removed hypothetically only where rescue has not already determined the endpoint. A different clinical question could justify a different combination.
Worked Example: Calculate the Result Under the Joint Rule
Worked Example: Calculate the Result Under the Joint Rule
Consider a small, entirely hypothetical dataset with five participants per group. All scores are on the 0–10 symptom-burden scale, and all participants have a week-24 endpoint. The listed score is the observed score unless the joint strategy calls for a composite or hypothetical value.
| Group | Participant history | Observed score | Value under specified strategy |
|---|---|---|---|
| New treatment | No listed event | 2 | 2 |
| New treatment | Discontinued assigned treatment | 4 | 4 |
| New treatment | Rescue medication | 8 | 10 |
| New treatment | Subsequent therapy, no rescue | 7 | 3 |
| New treatment | Rescue, then subsequent therapy | 9 | 10 |
| Control | No listed event | 3 | 3 |
| Control | Discontinued assigned treatment | 5 | 5 |
| Control | Rescue medication | 8 | 10 |
| Control | Subsequent therapy, no rescue | 8 | 7 |
| Control | Rescue, then subsequent therapy | 9 | 10 |
The hypothetical values 3 and 7 for participants who received subsequent therapy are illustrative outputs from a hypothetical-outcome model; they are not observed scores. In this example, no attempt is made to claim that a model could reliably estimate them from five participants per group.
For the new-treatment group, the joint-strategy mean is
For the control group, it is
The estimated difference, new treatment minus control, is \(5.8-7.0=-1.2\) points. In this hypothetical calculation, the new-treatment group has a mean symptom-burden score 1.2 points lower under the specified joint event strategies. This small-sample arithmetic illustrates the estimand definition; it is not a reliable estimate of a real treatment effect.
The table also shows why overlap rules matter. The participant with rescue followed by subsequent therapy receives a value of 10 under the stated precedence rule. Applying the hypothetical subsequent-therapy model to that participant instead would define a different joint strategy and could produce a different mean.
Worked Example: The Strategy Combination Changes the Question
Worked Example: The Strategy Combination Changes the Question
Use the same hypothetical participants, but now choose treatment policy for all three events. The endpoint is the observed week-24 score regardless of discontinuation, rescue, or subsequent therapy. The new-treatment mean is
and the control mean is
The treatment contrast is \(6.0-6.6=-0.6\) points, rather than \(-1.2\) points under the earlier joint strategy. Both calculations are internally consistent, but they answer different questions. Treatment policy includes outcomes after all three events as they occurred; the earlier strategy assigns a poor endpoint to rescue and uses hypothetical outcomes for certain subsequent-therapy histories. The numerical difference is not a correction for bias. It reflects different estimands.
Assumptions, Diagnostics, and Failure Modes
Randomization supports a comparison of assigned treatment conditions, but it does not automatically justify every event-specific strategy or identify every counterfactual outcome. A well-defined strategy requires that the event and endpoint definitions be sufficiently precise to assign a value to each relevant participant history. The outcome must also be collected at the planned time for treatment-policy strategies to use post-event outcomes.
Hypothetical strategies require extra care. Estimating an outcome under a condition that did not occur for some participants generally depends on assumptions about how observed information relates to those unobserved counterfactual outcomes. Analysts commonly need adequate measured predictors of event occurrence and outcome, sufficient overlap in the data for the modeled conditions, and a sufficiently appropriate outcome model. These assumptions are not guaranteed by randomization. Bayesian modeling represents uncertainty conditional on the chosen model and assumptions; it does not make unsupported counterfactual assumptions true.
Useful checks include tabulating each event separately and in combination, checking event timing and endpoint availability, and verifying that the strategy rule assigns one unambiguous endpoint interpretation to every history. For hypothetical components, inspect whether comparable participants and outcome information support the model predictions. Sparse or empty event-history groups, extreme predictions, and conclusions that change materially across reasonable modeling choices are warnings that the hypothetical component is weakly supported.
- Applying one strategy to every event by default. The correction is to justify each strategy against the clinical question, then state how they combine.
- Leaving overlaps unresolved. If rescue and subsequent therapy occur in sequence, state whether rescue determines the endpoint, both events are handled jointly, or another explicit rule applies.
- Confusing the strategy with a data-handling method. A hypothetical strategy defines a counterfactual target; deleting post-event observations is not, by itself, a valid way to estimate it.
- Treating an event as missing data. An intercurrent event affects the interpretation of the endpoint. Missing outcome data are a separate problem when a needed endpoint value is unavailable.
- Calling a strategy combination universally correct. The strategy is defensible only relative to a clearly stated clinical question and endpoint definition.
The principal alternative is not a single method but a different estimand: for instance, treatment policy for all events, a composite outcome for more than one event, or a joint hypothetical strategy that removes multiple events. Each alternative changes the target interpretation and may also change the assumptions or data needed for estimation. The limitation is fundamental: no statistical procedure can make two distinct clinical questions equivalent, and a precise strategy cannot compensate for poor measurement or unsupported counterfactual modeling.
Check Your Understanding
Answer each question using the event definitions and endpoint scale in the trial examples. Solutions follow the questions.
- If a participant discontinues assigned treatment but has no rescue or subsequent therapy, what endpoint value is used under the first joint strategy?
- Under the first joint strategy, what happens when rescue occurs before subsequent therapy?
- Why is the hypothetical subsequent-therapy value not simply the observed score after that therapy?
- Calculate the treatment-policy mean difference, new treatment minus control, from the worked example.
- Name one warning sign that the hypothetical component may be weakly supported.
Solutions. (1) Use the observed week-24 score, because discontinuation has a treatment-policy strategy. (2) Rescue takes precedence, so the composite endpoint is 10, even if subsequent therapy follows. (3) The strategy targets the score under a condition in which subsequent therapy did not occur; the observed post-therapy score is measured under a different treatment history. (4) The means are 6.0 and 6.6, so the difference is \(6.0-6.6=-0.6\) points. (5) Examples include sparse or empty event-history groups, extreme model predictions, or substantial sensitivity of conclusions to reasonable modeling choices.