What a Treatment-Policy Estimand Asks
An intercurrent event occurs after treatment assignment and affects either the interpretation or existence of subsequent outcomes. A participant might discontinue assigned treatment, switch treatments, use rescue medication, or experience another event that changes the course of care. A treatment-policy estimand asks about the outcome under each assigned treatment condition while retaining outcomes after such events. The question is therefore about the effect of assignment to treatment conditions as implemented in the trial, including the consequences of events and care that follow assignment.
The key is not to pretend that an intercurrent event did not happen. If a participant discontinues assigned treatment and then has an outcome measured at the planned time, that post-discontinuation outcome is part of the treatment-policy outcome. Similarly, outcomes after rescue medication or treatment switching remain included when those events are part of the specified treatment-policy question.
The definition requires the treatment conditions, target population, outcome variable, and assessment time to be specified. It also requires a clear account of what happens after an intercurrent event in each treatment condition: for example, whether usual care, rescue treatment, or treatment switching is allowed. Such details matter because they help define the outcomes \(Y_i(1)\) and \(Y_i(0)\). “Regardless of the event” does not mean that the event or subsequent care is irrelevant to the outcome.
For example, if the outcome is a symptom score at 24 weeks and lower scores are better, a treatment-policy contrast is the difference in mean 24-week scores under assignment to the two conditions, with 24-week scores after discontinuation or rescue medication included. A negative contrast then favors assignment to condition 1 on this scale. The contrast is not, by itself, the effect of continuously taking the assigned treatment or the effect among participants who adhered.
Worked Examples
Worked Example: A Mean Contrast With Post-Event Outcomes Included
Suppose a hypothetical randomized trial compares assignment to an active treatment with assignment to control. The outcome is a 24-week symptom score, in points, with lower scores indicating fewer symptoms. The 24-week score is measured even when participants discontinue their assigned treatment or start rescue medication. Among 80 participants assigned to active treatment, the mean score is 6.4 points; among 80 assigned to control, it is 7.1 points. These means include the measured post-event scores.
The estimated treatment-policy contrast, active treatment minus control, is $$ \widehat{\Delta}_{\mathrm{TP}}=\bar{Y}_1-\bar{Y}_0=6.4-7.1=-0.7\ \text{points}. $$
Thus, in this hypothetical trial, assignment to active treatment is associated with a mean 24-week symptom score 0.7 points lower than assignment to control. Because the outcome after an intercurrent event is retained, this summary targets the treatment-policy question, rather than a contrast restricted to participants who remained on their assigned treatment. This calculation alone does not give a standard error or confidence interval.
Worked Example: Combining Outcomes From Participants With and Without an Event
Consider another hypothetical trial with symptom scores measured after treatment discontinuation. Lower scores are better. In the active-assignment arm, 80% of participants have no discontinuation and have a mean score of 8 points; the remaining 20% discontinue and have a post-discontinuation mean score of 12 points. In the control-assignment arm, 90% have no discontinuation and have a mean score of 9 points; 10% discontinue and have a post-discontinuation mean score of 14 points. Assume the stated proportions and subgroup means describe all participants in each arm, so the outcomes are observed after the event as well as before it.
The full-arm mean for active assignment is the weighted average over event status: $$ \bar{Y}_1=0.80(8)+0.20(12)=6.4+2.4=8.8\ \text{points}. $$
For control assignment, it is $$ \bar{Y}_0=0.90(9)+0.10(14)=8.1+1.4=9.5\ \text{points}. $$
Therefore, $$ \widehat{\Delta}_{\mathrm{TP}}=8.8-9.5=-0.7\ \text{points}. $$
The contrast compares the full assigned groups, with each arm’s own event rate and post-event outcomes contributing to its mean. It is not the difference between the two “no discontinuation” means, which would be \(8-9=-1\) point and would answer a different question. The weighted calculation assumes that the subgroup proportions and means apply to the entire respective arm; if outcomes are missing within either subgroup, the displayed arithmetic cannot recover the full-arm means without additional assumptions or information.
Worked Example: Unequal Allocation Under Complete Randomization
Suppose a hypothetical trial has five participants, with three assigned to treatment 1 and two assigned to treatment 0 by complete randomization. For illustration, the fixed potential outcomes under each assignment are shown below. These numbers include any outcomes that would occur after an intercurrent event under the relevant assignment.
| Participant | \(Y_i(1)\), points | \(Y_i(0)\), points |
|---|---|---|
| 1 | 5 | 8 |
| 2 | 7 | 9 |
| 3 | 8 | 10 |
| 4 | 10 | 11 |
| 5 | 10 | 12 |
The finite-population mean under treatment 1 is \((5+7+8+10+10)/5=40/5=8\) points. Under treatment 0 it is \((8+9+10+11+12)/5=50/5=10\) points. Hence the finite-population treatment-policy contrast is \(8-10=-2\) points.
Each participant has probability \(3/5=0.6\) of assignment to treatment 1 and probability \(2/5=0.4\) of assignment to treatment 0. The probabilities are not equal because allocation is not 1:1. If participants 1, 2, and 4 happen to be assigned to treatment 1, the observed treatment-1 mean is \((5+7+10)/3=22/3=7.333\) points. Participants 3 and 5 are then assigned to treatment 0, giving an observed mean of \((10+12)/2=11\) points. The realized contrast is \(7.333-11=-3.667\) points, which differs from the finite-population contrast of \(-2\) points. Random assignment makes the arm-mean contrast unbiased over the randomization distribution; it does not require every realized allocation to reproduce the population contrast exactly.
Why the Randomized Arm-Mean Contrast Targets the Estimand
A short derivation makes the role of randomization precise. Suppose there are \(N\) participants, exactly \(n_a\) are assigned to condition \(a\), and assignment is complete randomization: every allocation with those arm sizes is equally likely. Let \(I_i(a)\) equal 1 if participant \(i\) is assigned to condition \(a\), and 0 otherwise. The observed outcome in that arm is \(Y_i(a)\) whenever \(I_i(a)=1\). If the outcome is observed for every participant, the arm mean can be written
Under this randomization scheme, each participant has the same probability \(n_a/N\) of assignment to condition \(a\). That probability is the same for each participant, but is not necessarily the same across arms. Taking the expectation over assignments, while holding the potential outcomes fixed, gives
Thus each arm mean is unbiased for the corresponding finite-population mean under that assignment. Subtracting the two equalities shows that \(\bar{Y}_1-\bar{Y}_0\) is unbiased for \(\Delta_{\mathrm{TP}}\). This result depends on randomization and on having the relevant outcomes observed; it does not establish that missing outcomes can be ignored. It also concerns the average contrast, not the unobservable pair of potential outcomes for any one participant.
For a target population described by a probability distribution rather than a fixed finite set, the corresponding interpretation is a difference in expected outcomes under the two assignment conditions. The trial’s observed arm means estimate that target contrast only to the extent that the trial participants and the analysis support the intended population inference.
Assumptions, Diagnostics, and Failure Modes
The estimand is a definition of the question; an estimator is a rule for using data to estimate it. A treatment-policy estimand can be defined even if outcomes after events are not successfully collected, but incomplete measurement creates an estimation problem. The observed-data mean then need not equal the full-arm mean simply because the missing values occurred after an intercurrent event. The distinction between an intercurrent event and a missing outcome is essential: the event is part of the clinical course, while missingness describes an outcome that was not observed.
A randomized arm-mean analysis has a straightforward design-based justification when assignment follows the stated randomization scheme, the outcome is defined consistently under both assignments, and outcomes are available for all randomized participants at the target time. For a causal interpretation, treatment conditions must be sufficiently well-defined that \(Y_i(a)\) has a coherent meaning for each participant. The usual potential-outcomes interpretation also assumes no interference between participants: one participant’s assignment does not change another participant’s outcome. If the treatment condition permits a range of post-assignment care, that care should be described, since it contributes to the treatment-policy outcome.
Useful checks include verifying that the outcome time and scale are the same in both arms, tabulating event frequencies by assigned arm, and checking whether outcome collection continued after each event. Compare the number randomized in each arm with the number contributing an outcome. If those counts differ, report the amount and pattern of missingness rather than describing the observed-case mean as automatically treatment-policy. Also check that the reported summary includes all observed post-event outcomes, not just outcomes obtained before discontinuation.
Common warning signs include censoring participants at discontinuation and then calling the resulting analysis treatment-policy, or excluding participants who used rescue medication. Both actions remove outcomes that the treatment-policy definition says to retain. Another warning sign is presenting an observed-case contrast without explaining how missing post-event outcomes were handled. If post-event outcomes are unavailable, recovering the defined estimand requires additional assumptions and an appropriate estimation strategy; changing the estimand silently is not a solution.
How This Question Differs From Other Strategies
The treatment-policy strategy is one way to specify how an intercurrent event enters an estimand. It differs from alternatives because it retains the outcome under the course of events and subsequent care that follows assignment, rather than redefining the outcome to remove or replace those consequences. The choice depends on the clinical question, not on which analysis is easiest to compute.
| Question being asked | How the event enters the outcome question | Contrast with treatment-policy |
|---|---|---|
| Treatment-policy | Use the outcome regardless of whether the event occurs. | Includes post-event outcomes and the consequences of subsequent care. |
| Hypothetical | Ask about outcomes under a specified scenario in which the event does not occur. | Targets a scenario different from the observed course after assignment. |
| Composite-strategy | Incorporate the event into the outcome definition. | Changes what is counted as the outcome rather than simply retaining the measured post-event outcome. |
| While-on-treatment | Focus on outcomes while participants are receiving assigned treatment. | Does not target the full post-event course in the same way. |
These contrasts are conceptual: each strategy requires a precise definition of the outcome and analysis. In particular, treatment-policy does not imply that an event has no effect, nor that its consequences should be adjusted away. It includes those consequences in the outcome under the assigned treatment condition.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A trial measures a 12-week outcome after participants stop assigned treatment. Should those measured outcomes be retained for a treatment-policy estimand? Explain.
- Under complete randomization, 120 participants are allocated with 72 to condition 1 and 48 to condition 0. What is each participant’s probability of assignment to each condition?
- In a symptom-score trial, the treatment-policy arm means are 5.8 and 6.3 points, with lower scores better. Calculate and interpret the contrast, condition 1 minus condition 0.
- Suppose 15% of participants in one arm have no outcome recorded after an intercurrent event. Does the treatment-policy definition alone tell the analyst to treat those outcomes as equal to the observed outcomes? Why or why not?
Solutions. 1. Yes. If the 12-week outcomes are measured, they are retained regardless of stopping assigned treatment. 2. Each participant has probability \(72/120=0.60\) of assignment to condition 1 and \(48/120=0.40\) of assignment to condition 0; the probabilities are equal across participants within an arm-assignment event, not across the two arms. 3. The contrast is \(5.8-6.3=-0.5\) points. Assignment to condition 1 has a mean score 0.5 points lower than assignment to condition 0 in the stated comparison. 4. No. The estimand specifies which outcome is wanted, not the value of an unobserved outcome or how to estimate it. Missing post-event outcomes require an estimation approach with assumptions appropriate to the missingness and data collection process.