Why Switching Changes the Question
An interim decision may lead a trial to offer a different treatment, or a participant may switch treatments under a prespecified rule. The resulting outcome reflects both the initial randomized assignment and what happened afterward. That does not make the outcome unusable; it means the treatment effect must be defined in a way that matches the question.
For example, a trial might ask whether assignment to Treatment B rather than Treatment A improves a 12-week symptom score under the trial’s switching policy. Alternatively, it might ask what the mean difference would be if participants remained on their initially assigned treatment throughout follow-up. A third question could compare two specified rules for switching. These are different causal questions, even if they use the same participants and endpoint.
The introductory material on rescue therapy and concomitant treatment distinguishes events that occur after randomization from treatment assignment itself. Here the central task is to state which treatment path the estimand concerns when switching follows an interim decision. The target population, endpoint, follow-up time, summary measure, and handling of switching all need to be clear.
Three Distinct Targets
A treatment-switching estimand needs to identify the treatment conditions being compared. The following notation makes the differences explicit. Write \(R\) for a specified trial policy, including how interim information leads to switching decisions. Write \(Y^{z,R}\) for the potential outcome under initial assignment \(z\) and that policy. The assignment effect under \(R\), for a mean outcome, is
This contrast includes whatever switching is part of \(R\). It is not the effect of receiving Treatment B rather than Treatment A continuously. If the protocol’s switching rule changes, then \(R\) changes; the assignment effect under the new policy need not equal the effect under the old one.
For a sustained-treatment question, write \(Y^{B\text{ throughout}}\) and \(Y^{A\text{ throughout}}\) for outcomes under treatment paths that maintain B or A, respectively, through the outcome time. The corresponding contrast is
This is a hypothetical comparison of sustained paths. Observed outcomes among participants who did not switch do not automatically equal these potential outcomes: remaining on treatment can depend on interim response, prognosis, or other post-randomization information.
For a strategy question, let \(g\) denote a rule that maps available history to treatment. For instance, one rule might switch participants with insufficient improvement at week 4 to B, while another keeps everyone on A. The contrast is
The rules need enough detail to be interpretable: when switching can occur, what information triggers it, and which treatment is given after the trigger. A vague condition such as “switch nonresponders” is incomplete unless response and the decision time are defined.
Worked Examples of Defining the Target
Worked Example: Naming an Assignment Effect
Suppose a hypothetical trial randomizes participants to A or B. At week 6, a prespecified interim decision permits participants assigned to A to switch to B if their symptoms have not improved sufficiently. The primary endpoint is a symptom score at week 12, in points, with lower scores better. What effect is targeted if the analysis compares the week-12 mean by initial randomized assignment and retains outcomes after switching?
The target is an assignment effect under the specified switching policy \(R\): participants are compared according to whether they were initially assigned A or B, and their later outcomes remain part of the outcome under that assignment-and-policy condition. In the notation above, the contrast is \(E[Y^{1,R}]-E[Y^{0,R}]\), in symptom-score points.
This is not a contrast between continuously receiving B and continuously receiving A. The A-assigned group can include outcomes after switching to B. To interpret the effect, the switching rule must be specified; changing who can switch or when can change the treatment condition represented by \(R\).
Worked Example: Calculating an Assignment Contrast With Switching
In a hypothetical randomized trial, the week-12 symptom scores are summarized as follows. Lower scores are better. Within each initial-assignment group, the reported means include outcomes after switching.
| Initial assignment | Number of participants | Mean week-12 score |
|---|---|---|
| A | 100 | 9.4 |
| B | 100 | 6.2 |
Define the assignment contrast as B minus A. The estimated difference in these hypothetical group means is
Thus, under the trial’s specified policy, the mean score was 3.2 points lower among participants initially assigned B than among those initially assigned A. This is a descriptive calculation of the assignment contrast in the hypothetical data; it does not establish that sustained B causes a 3.2-point reduction. Nor does the result imply that every participant benefited.
If the switching rule were altered, the outcomes represented by the assignment-and-policy conditions could also change. Therefore the policy belongs in the interpretation of the contrast, not just in an operational note.
Worked Example: Contrasting Two Switching Strategies
Consider a hypothetical trial asking whether a rule that switches participants with inadequate improvement at week 4 to B performs better than a rule that keeps everyone on A. Suppose the week-12 mean scores under these two fully specified strategies are 7.1 points and 8.0 points, respectively. Lower scores are better.
The strategy contrast, rule 1 minus rule 0, is
Under these hypothetical values, the rule that switches participants meeting the week-4 criterion has a mean score 0.9 points lower than the rule that keeps everyone on A. This comparison concerns the two rules as whole treatment strategies. It does not isolate the effect of B among those who switched; participants following the rules may differ in their treatment paths and in the histories that trigger those paths.
Why the Assignment Effect Is a Policy-Specific Mixture
A simple decomposition shows why the assignment contrast can differ from a sustained-treatment contrast. Suppose that, under initial assignment \(z\), participants either switch or do not switch. Let \(S^z=1\) indicate switching under that assignment and policy. The law of total expectation gives
Each assignment mean is therefore a mixture of outcomes among those who switch and those who do not, weighted by the switching frequencies under that assignment. Subtracting the two assignment means compares two mixtures, which may have different weights and different participant histories. It does not simply subtract the mean outcome under sustained B from the mean outcome under sustained A.
This identity is a description of the assignment-policy outcome distribution, not a method for recovering the sustained-treatment effect. In particular, conditioning on observed switching status can select participants based on post-randomization information. A comparison of switchers with non-switchers can therefore mix treatment effects with differences in prognosis or interim response.
Assumptions and Specification Checks
The potential-outcome definitions are meaningful only when the treatment conditions are sufficiently well specified. For an assignment effect under \(R\), define which assignments are possible, how the interim rule operates, which treatments can follow, and when the outcome is measured. For a sustained-treatment estimand, state what “throughout” means, including any allowed interruptions or discontinuations. For a strategy estimand, specify the decision times, eligibility to switch, triggering information, and treatment after each possible decision.
Causal interpretation also relies on standard conditions. Random assignment supports comparability of the initial assignment groups, subject to the trial’s design and implementation. Consistency requires that the observed outcome for a participant following a specified condition agrees with the potential outcome for that condition. The treatment conditions must not hide materially different versions that would make the same label ambiguous. The usual no-interference condition means one participant’s outcome is not affected by another participant’s assignment or switching, except insofar as the defined trial condition accounts for such effects.
For assignment effects, outcomes must be interpreted under the policy actually specified, and outcome ascertainment must not silently exclude switchers. For sustained-treatment or strategy targets, defining the target does not by itself ensure that the trial data can identify or estimate it. The issue of how an estimator or test relates to an estimand is a separate question.
Diagnostics, Alternatives, and Limitations
A practical diagnostic is to write the treatment condition in ordinary language and ask whether two participants who follow it in different ways would still count as receiving the same condition. If the answer is unclear, the condition needs more detail. Also check protocol versions: an interim decision that changes who may switch, the trigger, or the timing can define a different policy condition.
| Target | Question answered | What must be specified |
|---|---|---|
| Assignment effect | What is the effect of initial assignment under the trial policy? | Initial assignments and the switching policy under which outcomes are included |
| Sustained-treatment effect | What would the contrast be if each treatment were maintained? | The sustained treatment paths and what counts as interruption or discontinuation |
| Switching-strategy effect | What is the effect of following one switching rule rather than another? | Decision times, triggers, available treatments, and actions under each rule |
The assignment effect is often the clearest target when the practical question concerns the consequences of choosing an initial assignment under the trial’s policy. A sustained-treatment effect answers a different, often hypothetical question about maintaining treatment. A switching-strategy effect is useful when the decision rule itself is the clinical or policy choice. These targets are alternatives, not interchangeable labels for the same contrast.
A common mistake is to exclude switchers and compare the remaining participants as though they represented the sustained-treatment groups. That reasoning is generally wrong because switching can depend on post-randomization response or prognosis. The correction is to state the intended treatment condition first. If the target is assignment under a policy, include outcomes after switching in that policy-specific target. If the target is sustained treatment or a switching strategy, define that counterfactual path or rule explicitly and do not equate it with a simple observed subgroup comparison.
The main limitation is that an estimand specifies the question, not whether available data can answer it without additional assumptions. A well-defined sustained-treatment or strategy contrast may be difficult to estimate when treatment histories and decision factors are not adequately observed. Also, an effect under one switching policy should not be assumed to transport unchanged to a materially different policy.
Check Your Understanding
For each question, identify the relevant target or calculate the requested contrast.
- A randomized trial compares week-12 outcomes by initial assignment and includes outcomes after protocol-permitted switching. What kind of effect is this, and what policy detail must accompany its interpretation?
- Hypothetical mean scores are 5.8 points under strategy \(g_1\) and 6.5 points under strategy \(g_0\), with lower scores better. Calculate the contrast \(g_1-g_0\) and interpret it.
- Why does comparing observed switchers with observed non-switchers not generally estimate the effect of sustained treatment?
- A protocol changes the week-4 switching trigger to a week-6 trigger. Can an assignment effect under the original switching policy automatically be interpreted as the effect under the revised policy? Explain.
Solutions. (1) It is an assignment effect under the trial’s switching policy. The interpretation must state the policy, including when and under what conditions switching is allowed. (2) \(5.8-6.5=-0.7\) points. Under the hypothetical values, the mean score under \(g_1\) is 0.7 points lower than under \(g_0\). (3) Switching may depend on post-randomization response or prognosis, so the two observed groups can differ for reasons other than treatment path. Their observed means do not automatically represent sustained-treatment potential outcomes. (4) No. The trigger and timing define part of the policy, so the revised protocol represents a different switching condition; the original effect cannot automatically be assigned the same interpretation under the revision.