What a While-on-Treatment Contrast Describes
A clinical outcome measured during treatment answers a different question from the same outcome measured regardless of treatment continuation. If a participant stops assigned treatment, switches therapy, or begins rescue treatment, follow-up may still produce valuable data. A while-on-treatment analysis deliberately uses outcomes only up to a specified event and does not use later outcomes for its target summary. Its clinical meaning therefore depends on which event ends the relevant period, when that event occurs, and how the pre-event outcomes are summarized.
The introductory material on intercurrent events distinguishes events that affect the interpretation of an outcome from missing observations. Here the key interpretation is narrower: a while-on-treatment strategy uses data collected before the specified event. It does not, by itself, define a single universal endpoint or imply that people who remain on treatment form comparable groups.
For clarity, let \(A\) denote randomized treatment, with \(a\) indexing a treatment group. Let \(E\) be the time of the specified event, such as permanent discontinuation of assigned treatment, and let \(Y(t)\) be the outcome at time \(t\). A fixed-time on-treatment summary might compare the mean outcome at time \(t\) among participants who have not yet had the event. An alternative summary might average each participant’s available pre-event measurements and then compare those participant-level averages. These summaries are not interchangeable: the first concerns a particular time among those still on treatment then; the second concerns an average over potentially different periods.
This formula illustrates a fixed-time, group-specific on-treatment comparison. The condition \(E>t\) selects participants whose specified event has not occurred by time \(t\). If higher outcomes are better, a positive contrast favors group 1 on this scale. It is not automatically the treatment effect in the full randomized population at time \(t\), nor is it automatically the effect among a common group of people who would remain on treatment under either assignment.
Specify the Event and the Outcome Summary
“While on treatment” is not sufficiently precise for clinical interpretation until the event rule is operational. For example, does the period end at the first missed dose, permanent discontinuation, a switch to another drug, or initiation of rescue therapy? If temporary interruptions are allowed, does the period end at the interruption or only after a prespecified duration? The definition should also state whether the event time is the date it occurs, the next scheduled assessment, or another protocol-defined time.
The outcome summary must match that rule. At a fixed visit, the analysis could use outcomes at that visit only for participants still on treatment. For a longitudinal summary, it could use all scheduled measurements before the event, with a stated weighting rule. Averaging all available records without specifying whether participants or visits receive equal weight can silently change the target: participants with longer event-free follow-up may contribute more records and thus more influence.
Other assumptions concern how the data represent this target. The treatment groups should be defined by the intended treatment assignment or exposure rule, and measurements before the event must be collected and recorded according to the stated schedule. If some pre-event outcomes are missing, excluding them does not automatically preserve the intended summary; assumptions or methods for missing data are needed. This is distinct from intentionally not using outcomes after the event.
Worked Example: A Fixed-Time Contrast Among Participants Still on Treatment
In a hypothetical randomized trial, six participants per group are assigned treatment A or B. The outcome is a symptom score at week 12, with higher scores indicating worse symptoms. Two participants in each group discontinued assigned treatment before week 12. The week-12 scores for those who remained on treatment are A: 8, 7, 6, 5 and B: 7, 6, 6, 5. Assume these scores are recorded for every participant who remained on treatment, and define the event as discontinuation.
The group A mean among participants still on treatment is
The corresponding group B mean is
The observed fixed-time contrast, A minus B, is \(6.50-6.00=0.50\) symptom-score points. Under the stated summary, the week-12 mean among participants who had not discontinued was 0.50 points higher (worse) in group A than in group B. The calculation uses only week-12 outcomes before the event for participants still on treatment; it does not assign a week-12 score to the discontinuers. It also does not establish a 0.50-point treatment effect for all randomized participants.
Randomization supports comparability of the groups at assignment, but it does not ensure that the participants still on treatment at week 12 remain comparable. If discontinuation is related to symptoms or prognosis, the two on-treatment subsets can differ in clinically relevant ways.
Why Event-Related Follow-Up Changes Interpretation
A while-on-treatment contrast at a later visit conditions on an event that can occur after randomization. Consequently, the membership of each group’s on-treatment subset can depend on treatment assignment and on participants’ characteristics. Even when the original groups were randomized, the subsets remaining at time \(t\) need not be balanced.
The reason can be seen by writing the conditional mean as an average over a prognostic characteristic \(U\), such as baseline disease severity. The weights in that average are the proportions of each prognostic group among participants who remain on treatment. If those proportions differ by treatment group, the conditional means can differ even when outcomes are identical between treatments within each prognostic group.
Thus, interpretation depends both on within-group outcome levels and on the selected groups’ composition. Event-related follow-up is clinically informative: rates and timing of discontinuation may reflect tolerability, lack of efficacy, or other causes. A contrast that excludes post-event outcomes should be read alongside that information, not as though the event had no clinical consequences.
Worked Example: Different Risk Profiles Can Create a Contrast
Consider a hypothetical study with equal numbers of participants with good and poor baseline prognosis in each randomized group. Suppose the week-8 symptom score is 8 for everyone with good prognosis and 4 for everyone with poor prognosis, regardless of assigned treatment. There is no treatment difference within either prognosis group. By week 8, however, the on-treatment subset in group A contains 8 participants with good prognosis and 2 with poor prognosis, while the subset in group B contains 2 with good prognosis and 8 with poor prognosis. Assume all these week-8 scores are observed and the event is discontinuation.
For group A, the fixed-time mean among those still on treatment is
For group B, it is
The observed contrast is \(7.20-4.80=2.40\) points, favoring group B because lower symptom scores are better. Yet within each prognosis category the score is identical across treatments. The difference in these selected means arises entirely from the on-treatment subsets’ different prognosis profiles. This calculation shows why randomization alone does not license a causal interpretation of a comparison conditional on remaining on treatment.
Summaries Over Different Pre-Event Periods
A longitudinal while-on-treatment summary can use measurements before the event rather than a single fixed visit. The weighting rule determines its meaning. One transparent option is to calculate a mean of each participant’s available scheduled pre-event outcomes and then average those participant-level means within each randomized group. This gives each participant equal weight, even if their event-free follow-up lengths differ. It is still an average over periods of varying duration, not an average at a shared follow-up time.
Worked Example: Equal-Participant Weighting of Pre-Event Measurements
In a hypothetical trial, measurements are scheduled at weeks 4, 8, and 12. Scores are recorded only before discontinuation. Group A has three participants with pre-event score sequences \(10,8\); \(8,6,4\); and \(7,5\). Group B has sequences \(9,7\); \(7,5,3\); and \(6,4\). Assume all scheduled pre-event scores are recorded. Define each participant’s summary as the arithmetic mean of their observed pre-event scores, then give each participant equal weight.
The three group A participant means are \((10+8)/2=9\), \((8+6+4)/3=6\), and \((7+5)/2=6\). The group A mean of participant-level summaries is
The group B participant means are \((9+7)/2=8\), \((7+5+3)/3=5\), and \((6+4)/2=5\). Therefore
The stated summary is one point higher in group A, averaged first within participant over observed pre-event visits and then equally across participants. It is not a week-12 contrast: some participants contribute only weeks 4 and 8, while others contribute all three visits. A visit-weighted average, or a time-weighted area under the outcome curve, would answer a different question and could produce a different result.
Comparison, Diagnostics, and Limitations
A while-on-treatment strategy differs from other common ways of handling an intercurrent event. The alternatives answer different clinical questions; none is automatically preferable without specifying the objective.
| Approach | How post-event outcomes are treated | Interpretive distinction |
|---|---|---|
| While-on-treatment | Excluded from the specified outcome summary | Describes an outcome before the event or among those still event-free at a stated time. |
| Treatment-policy | Included regardless of the event | Describes outcomes under treatment assignment while allowing subsequent events and care to occur. |
| Hypothetical | Estimated as if the event had not occurred | Requires assumptions to identify outcomes under that counterfactual scenario. |
| Composite | Represented through a component of a combined outcome | Changes the outcome definition so the event itself contributes to the summary. |
| Principal-stratum | Targets a subgroup defined by event status under treatment alternatives | Concerns a latent subgroup, such as people who would remain on treatment under either assignment. |
The distinction from a principal-stratum interpretation is especially important. A mean among observed participants still on treatment in group A and a mean among observed participants still on treatment in group B generally concern different selected sets. They do not identify the contrast among people who would remain on treatment under either assignment. The introductory material on principal-stratum interpretation and identification addresses that separate target.
Practical checks help reveal when the intended interpretation is strained:
- Tabulate event frequency and timing by group. A large difference in discontinuation or a shift in event timing changes which outcomes can contribute.
- Describe reasons for events and baseline characteristics. Differences in reasons or prognostic profiles among those remaining can indicate selection that matters clinically.
- Check assessment timing and pre-event data completeness. Missing scheduled visits before discontinuation are not handled merely by excluding post-event observations.
- Verify the event rule and summary weighting. Ambiguous definitions, unequal record weighting, or measurements collected after the defined event can make the reported contrast difficult to interpret.
A common mistake is to say, “The analysis excludes outcomes after discontinuation, so it estimates the treatment effect among people who would adhere.” The correction is that exclusion defines which outcomes enter the chosen summary; it does not identify who would adhere under both assignments. In a fixed-time analysis, each arm’s observed on-treatment subset can be selected differently. A further limitation is that a summary over pre-event measurements can mix different follow-up durations, so its clinical meaning depends on the specified visit schedule and weighting rule. The contrast should therefore be reported with the event definition, event patterns, and outcome-summary rule.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- At week 16, treatment groups are compared using outcomes only for participants who have not discontinued assigned treatment. What population does each mean directly describe?
- In a longitudinal pre-event summary, why must the analyst state whether participants or visits receive equal weight?
- Suppose discontinuation is more frequent among participants with worsening symptoms in one group. Why might the later on-treatment contrast be difficult to interpret as a full-population treatment effect?
- What is the difference between excluding post-discontinuation outcomes and estimating outcomes under a hypothetical no-discontinuation scenario?
Solutions. (1) Each mean describes participants in that treatment group who remain event-free at week 16; the two means may concern different selected subsets. (2) Participants can contribute different numbers of pre-event visits. Equal participant weighting gives each participant’s pre-event average the same influence, whereas equal visit weighting gives more influence to participants with more recorded visits. These are different summaries. (3) The event can select participants according to outcome prognosis, changing the composition of the on-treatment subset. Randomization at assignment does not guarantee comparability after conditioning on event-free status. (4) Excluding post-event outcomes defines a summary based on pre-event data. A hypothetical strategy instead targets outcomes that would occur in a scenario where discontinuation did not occur and requires assumptions to estimate those unobserved outcomes.