Rescue Therapy and Concomitant Treatment

Why Rescue and Concomitant Therapy Matter

Rescue medication may relieve symptoms when assigned treatment is insufficient. Other concomitant therapy may be permitted, restricted, or adjusted during a trial. These choices are not merely operational: they help define what treatment effect the trial is intended to estimate. If the rescue rules change during an adaptive trial, the same comparison of randomized groups may represent a different clinical question before and after the change.

The key distinction is whether the target includes outcomes after rescue or concomitant treatment, or instead asks what outcomes would have been under a hypothetical scenario in which that therapy was not used. The first is often directly aligned with a pragmatic treatment policy. The second can answer a useful scientific question, but requires assumptions about outcomes that were not observed.

For a simple two-arm comparison, let \(Y^{a,\pi}\) denote the outcome that would be observed under assignment \(a\) and a specified rescue or concomitant-therapy policy \(\pi\), where \(a=1\) denotes the experimental assignment and \(a=0\) the control assignment. For an outcome where higher values are better, a treatment-policy contrast under policy \(\pi\) is

The superscript \(\pi\) matters: the contrast concerns assignment under that policy, not necessarily a drug effect isolated from all subsequent care. A hypothetical no-rescue contrast can be written \(E[Y^{1,\text{no rescue}}]-E[Y^{0,\text{no rescue}}]\). It asks a different question. Specifying the outcome scale, follow-up time, population, and policy is essential for both.

Treatment-Policy and Hypothetical Targets

Under a treatment-policy target, participants remain part of the comparison after rescue use, treatment discontinuation, or use of permitted concomitant medication. Their outcomes reflect the consequences of assignment together with what happened under the stated care policy. For example, if rescue is available when symptoms worsen, its availability and use are part of the treatment strategy being compared. This can be a clinically relevant question: what happens when patients are assigned to each study treatment in a setting where rescue is provided according to these rules?

A hypothetical target instead asks what the outcomes would have been if the specified additional therapy had not occurred. Since a rescued participant’s outcome without rescue is unobserved, the target cannot generally be obtained by simply removing rescued participants. Estimation requires a model or other justified method to predict their outcomes under the hypothetical condition. The assumptions must be plausible given the recorded data and clinical setting.

Concomitant therapy needs the same precision. A target might compare assignment under a policy that permits stable background therapy, or it might ask about outcomes if a particular co-treatment had not been initiated. “Adjusting for concomitant therapy” is not, by itself, a definition of either target. The target depends on which therapies are allowed, when they can be used, and whether their effects are included or removed from the question.

A treatment-policy estimand is not synonymous with an analysis that mechanically includes every observed participant. The analysis must still address missing outcomes and follow the specified outcome and population definitions. Nor does a hypothetical target become valid merely because a statistical model includes rescue use as a covariate. The estimand states the question; the estimator is the procedure used to estimate its answer.

Why Excluding Rescue Users Does Not Answer the Hypothetical Question

Rescue use usually follows participants’ evolving health. Those with worse symptoms may be more likely to receive it, and assigned treatment may itself affect whether rescue is needed. Consequently, the participants who do not receive rescue are not generally a randomized comparison. Comparing only those participants conditions on a post-randomization event and can select groups with different prognoses.

A hypothetical no-rescue estimand instead concerns each randomized group under the condition that rescue did not occur. To estimate it, analysts must predict the missing no-rescue outcomes for people who were rescued. One possible approach models outcomes using recorded baseline and time-varying information before rescue, then predicts outcomes under no rescue. Such a procedure needs assumptions; the model does not make the unobserved outcomes known.

Protocol Changes Can Change the Policy in the Estimand

Suppose an interim review leads to a protocol amendment that changes the rescue threshold, permitted concomitant medication, or timing of rescue. If outcomes are still collected and analyzed regardless of rescue, the treatment-policy estimand under the original policy has not automatically been preserved. Participants enrolled under the revised rules experience a different policy, and their treatment outcomes may differ for that reason.

If the policy change is part of the intended clinical strategy, the target can be defined as assignment under the revised policy. If the intention is to retain a fixed policy target, data from participants treated under a changed policy may not directly estimate it without additional assumptions or a justified adjustment. Pooling stages without describing the policy can yield a contrast averaged across multiple policies, rather than a contrast under one clearly stated policy.

Justification: Why the Pooled Contrast Is Policy-Weighted

The pooled result above follows directly from averaging outcomes across the stage-specific policies. Let \(S\) identify the stage, and suppose the same stage weights \(w_s\) apply to both randomized groups, with \(\sum_s w_s=1\). Then the pooled group mean is the weighted sum of its stage-specific means. Subtracting the pooled control mean from the pooled experimental mean gives:

Thus the pooled contrast averages stage-specific contrasts under policies \(\pi_s\), using the stage weights. It equals a contrast under one common policy only if the policies are in fact the same, or if further assumptions justify treating the policy-specific contrasts as equivalent for the target in question. If stage weights differ between treatment groups, even this simple weighted-contrast interpretation need not hold; the pooled difference can also reflect different stage composition.

Diagnostics, Alternatives, and Limitations

A protocol review should track rescue and concomitant therapy as components of the treatment condition, not only as background details. Useful checks include:

• Compare the rescue criteria, permitted therapies, and implementation dates across stages or sites.
• Summarize the proportion receiving rescue, time to first rescue, and reasons for use by randomized group and stage.
• Check whether relevant symptom history and other predictors of rescue and outcome were recorded before rescue use.
• Inspect overlap: if some histories almost always lead to rescue, a no-rescue model may require unsupported extrapolation.
• Report missing outcomes and whether their occurrence is associated with rescue use or protocol period.

These checks do not establish that a model is correct. Large differences in rescue frequency across stages, sparse observations for particular clinical histories, or poorly recorded concomitant therapy are warnings that the target may be difficult to estimate or that the policy-specific interpretation needs to be explicit.

The main alternatives answer different questions. A treatment-policy analysis retains outcomes after rescue and describes assignment under the rescue rules actually specified. A hypothetical analysis targets outcomes under a stated no-rescue or no-co-treatment condition, with additional modeling assumptions. A composite strategy can incorporate rescue into the outcome definition, while a while-on-treatment approach focuses on outcomes before a specified event; neither should be substituted without stating the changed target. A principal-stratum comparison concerns a defined subgroup based on potential post-randomization behavior and does not answer the overall treatment-policy question.

The treatment-policy approach is often easier to interpret operationally, but it can be sensitive to how the rescue policy is implemented and to changes in access to additional therapy. The hypothetical approach can clarify a scientific question about treatment without rescue, but its answer can be weakly supported when rescue is common or its determinants are poorly measured. A medically necessary rescue therapy should not be withheld simply to simplify estimation.

Common Mistake

Wrong reasoning: “Rescue medication obscures the study drug effect, so I will exclude everyone who received rescue. The remaining participants estimate the effect without rescue.” The correction is that rescue users are selected after randomization, often because their health changed. Excluding them compares selected groups and does not generally recover the outcomes all participants would have had without rescue. A hypothetical target must be stated explicitly and estimated using assumptions that support prediction of those unobserved outcomes.