Why Switching Changes the Question
A participant assigned to a control treatment may later switch to the experimental treatment, or a participant may begin another therapy after trial treatment is stopped. These events matter because they can change what treatment is received and therefore what outcome is observed. But they do not, by themselves, determine which treatment effect a trial should estimate. That depends on the clinical question.
If the question is about the consequences of assignment to a treatment strategy in a setting where switching and later therapies may occur, outcomes after those events belong in the comparison. If the question is about what would happen if participants stayed on their assigned treatment and did not begin specified later therapies, the target is different: it concerns a hypothetical treatment condition. The two contrasts can differ even when calculated from the same trial.
An operational definition should state which treatment changes count, when they count, and how they are ascertained. For example, a switch might mean starting the experimental drug while assigned to control, whereas subsequent therapy might mean starting any specified nontrial drug after stopping trial treatment. The definitions need not be the same. A brief interruption, a dose change, and a switch to another treatment are not interchangeable event definitions.
State the Treatment Conditions Before Comparing Outcomes
Let \(A\) denote randomized assignment, with \(A=1\) for the experimental strategy and \(A=0\) for the control strategy. Let \(Y\) be the outcome at a specified time, with larger values defined here as better. A treatment-policy question compares outcomes according to assignment while allowing switching and subsequent therapy to occur as they would in the trial setting. A hypothetical question instead compares outcomes under specified conditions that rule out one or more of those events.
These conditions must be explicit. “Effect of treatment” is too vague: it could mean an effect of assignment with later treatment changes included, or an effect under continued assigned treatment without switching. A hypothetical condition also needs a time window and an event definition. For instance, “no switch through week 24” is more precise than “no switching.”
The earlier material on treatment-policy estimands and hypothetical estimands establishes these strategies in general. Here the important application is that switching and subsequent therapy need not be handled identically. A trial may retain outcomes after switching for its treatment-policy question, while defining a separate hypothetical question that rules out switching but allows other subsequent therapy. Or it may define a hypothetical condition that rules out both. The estimand must say which.
Worked Example: Classifying a Treatment Change
Worked Example: Classifying a Treatment Change
In a hypothetical 24-week trial, participants are assigned to an experimental drug or control. At week 10, a participant assigned to control starts the experimental drug. Another participant stops trial treatment at week 12 and begins a specified outside therapy at week 14. Classify the events for two proposed questions.
Question 1: What is the effect of assignment in the trial setting? The week-10 change is switching, and the outside therapy is subsequent therapy. Under a treatment-policy strategy for both events, the week-24 outcome is interpreted under the participant’s assigned group and includes the course of treatment actually received.
Question 2: What would the effect be if no one switched from assigned treatment? The week-10 change is the event ruled out by the hypothetical condition. The outside therapy at week 14 is not automatically ruled out: it is excluded only if the question also specifies no such subsequent therapy. Thus the event classification is determined by the operational definitions, and the strategy is determined by the question.
This classification alone does not calculate either treatment effect. It prevents a common ambiguity: treating every post-randomization treatment change as if it had the same role in every estimand.
Treatment-Policy Contrast: Keep the Post-Event Outcomes
For a treatment-policy question, outcomes after switching or subsequent therapy are part of the outcome under the assigned strategy. If the outcome is observed for every randomized participant, a simple estimate is the difference between the mean outcome in the two assigned groups. This is an assignment effect in the trial’s treatment setting, not an effect of actually receiving only the assigned drug.
Worked Example: A Treatment-Policy Mean Difference
Suppose a hypothetical randomized trial assigns 100 participants to each group. The outcome is improvement in a symptom score at week 24, in points, with larger improvement better. The experimental-assignment group has a mean improvement of 6.2 points; this group includes participants who received subsequent therapy. The control-assignment group has a mean improvement of 4.8 points; this group includes 15 participants who switched to the experimental treatment and 10 who began subsequent therapy. Assume the week-24 outcome is recorded for all 200 participants.
For the treatment-policy contrast, retain each participant in the group to which they were assigned and use the outcomes after the treatment changes:
The estimated mean improvement is 1.4 points greater under assignment to the experimental strategy than under assignment to control in this trial setting. This result does not say that receiving the experimental drug itself caused a 1.4-point difference: the control-group mean includes outcomes after switching and subsequent therapy.
Randomization supports the assignment comparison when assignment is implemented as planned, the outcome is defined and observed for the participants being compared, and the comparison preserves randomized groups. In potential-outcome terms, let \(Y^1\) and \(Y^0\) be each participant’s outcomes under assignment to the two strategies, including the treatment changes that would follow each assignment in the trial setting. Randomization and consistency give the justification:
The equality follows because randomization makes \(A\) independent of \((Y^1,Y^0)\), and consistency means the observed outcome equals \(Y^a\) for a participant assigned \(A=a\). The right-hand side is an assignment contrast under the specified setting. If outcomes are missing, the observed group means alone need not estimate this contrast; the missing outcomes require appropriate handling for the chosen estimand.
Hypothetical No-Switch Contrast: Outcomes Under a Different Condition
For a hypothetical no-switch question, outcomes observed after switching do not directly reveal what those same participants would have experienced had they remained on assigned treatment. That counterfactual outcome is not observed for a person who switched. Estimating the hypothetical contrast therefore requires assumptions and an estimation approach appropriate to the data. In particular, simply dropping switchers changes the groups being compared and generally does not reproduce randomization.
A credible analysis needs the hypothetical condition to be sufficiently well specified, adequate information about participants’ treatment histories and reasons for switching, and a defensible basis for predicting outcomes under the no-switch condition. Depending on the approach, this can require that relevant predictors of switching and outcome are measured, that there are comparable participants who followed the treatment condition in relevant histories, and that the model used to predict counterfactual outcomes is adequate. These are substantive assumptions, not consequences of randomization alone.
Worked Example: A Hypothetical No-Switch Contrast
Continue with a hypothetical 24-week trial with 100 participants assigned to each group. Suppose 20 control-assigned participants switched to the experimental treatment. The 80 control-assigned participants who did not switch had a mean improvement of 5.0 points. The observed switchers had a mean improvement of 4.0 points after switching. The observed control-assignment mean is therefore consistent with
The 4.0-point mean for switchers is an outcome under switching, not their unobserved outcome under no switching. Suppose, for illustration, that a specified prediction model estimates that these 20 switchers would have averaged 5.5 points had they stayed on control. Holding the observed non-switcher outcomes fixed for this calculation gives a hypothetical control mean of
Suppose the same model-based analysis estimates a hypothetical no-switch mean of 6.0 points for the experimental-assignment group. Then the hypothetical contrast is \(6.0-5.1=0.9\) points. This is not the treatment-policy estimate of 1.4 points. It relies on the model’s counterfactual predictions; the arithmetic does not validate the assumptions that produced 5.5 and 6.0.
A sensitivity check can vary the predicted mean for control switchers. If it were 4.5 rather than 5.5 points, the hypothetical control mean would be \(\{80(5.0)+20(4.5)\}/100=4.9\) points and the contrast, retaining 6.0 for the experimental group, would be 1.1 points. Reporting how the contrast changes under plausible predictions makes the dependence on unobserved outcomes visible.
Choose Among Strategies, Then Check the Data
Treatment policy and hypothetical strategies answer different clinical questions; neither is universally preferable. Other strategies already introduced in the estimand framework may also be appropriate when they match the question. For example, a composite strategy incorporates the event into the outcome definition, while a principal-stratum strategy targets a defined subgroup. Those alternatives should not be substituted silently for either an assignment effect or a no-switch effect.
| Strategy for switching or subsequent therapy | What the contrast means | Key caution |
|---|---|---|
| Treatment policy | Outcome under assignment, including outcomes after the event | Interpret as an assignment effect in the specified care setting |
| Hypothetical | Outcome under assignment if the specified event had not occurred | Requires assumptions to estimate outcomes that are counterfactual for event participants |
| Composite | Effect when the event is incorporated into the outcome definition | The composite outcome must be clinically interpretable |
| Principal stratum | Effect in a subgroup defined by potential event behavior | Membership may not be directly observable for each participant |
Useful diagnostics start with the event record. Check whether the timing and type of each switch or subsequent therapy are captured consistently, and whether the definition used in analysis matches the question. For a hypothetical analysis, examine how many participants follow the no-switch condition within relevant histories and whether model predictions rely heavily on extrapolation. Compare results across reasonable prediction assumptions. For a treatment-policy analysis, confirm that post-event outcomes are included as intended and that missing outcomes have not been mistaken for the intercurrent event itself.
A major failure mode is to compare only participants who remained on assigned treatment. Switching may be related to prognosis, symptoms, or treatment response. Conditioning on observed switching can therefore produce groups that differ in ways randomization did not balance. Another failure is to label a treatment-policy estimate as an effect of continued treatment: its outcomes include post-switch experience, so that interpretation does not follow.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In a randomized trial, a control-assigned participant starts the experimental drug at week 8. What is the event, and what happens to the participant’s week-24 outcome under a treatment-policy strategy?
- A hypothetical trial has assignment-group means of 7.1 and 5.6 points, with higher values better. Calculate and interpret the treatment-policy difference.
- Why is the mean among participants who did not switch not automatically an estimate of the hypothetical no-switch mean?
- In the worked example, recalculate the hypothetical control mean if the 20 switchers’ predicted no-switch mean is 6.0 points rather than 5.5. If the experimental hypothetical mean stays 6.0, calculate the contrast.
Solutions. (1) It is treatment switching. The participant remains in the control-assignment group, and the week-24 outcome after switching is included in that group’s treatment-policy outcome. (2) The difference is \(7.1-5.6=1.5\) points. The estimated mean outcome is 1.5 points higher under experimental assignment in the specified trial setting. (3) Switching may depend on prognosis, response, or other factors, so non-switchers can be a selected group; their outcomes alone do not supply switchers’ counterfactual outcomes under no switching. (4) The hypothetical control mean is \(\{80(5.0)+20(6.0)\}/100=(400+120)/100=5.2\) points. The contrast is \(6.0-5.2=0.8\) points, conditional on the stated prediction assumption.