Why a Treatment Rule Is a Treatment Condition
Some treatment strategies are defined by a single choice at baseline. Others specify what to do as a patient’s clinical history unfolds: continue treatment after a favorable response, switch after inadequate response, or add rescue treatment after a prespecified event. In the latter case, the intervention is not just a drug name or initial assignment. It is a rule linking patient history to treatment.
This distinction matters in an adaptive trial. A protocol might revise a response threshold, change which treatment follows a particular response, or replace one decision rule with another. Even if the trial continues to use the same outcome and target population, a changed rule can define a different treatment condition and therefore a different estimand. The earlier material on treatment conditions as fixed intervention strategies provides the starting point: a treatment condition must specify the intervention being compared, not merely label a trial arm.
A history may include baseline characteristics, prior treatment, response measurements, and intercurrent events observed before the decision. The rule must be precise about which information is used and when it is available. For example, “switch nonresponders” is incomplete if the response measure, assessment time, and threshold for nonresponse are unspecified.
A dynamic regime is not the same as the sequence of treatments a particular patient happens to receive. The regime is the intervention rule; the patient’s realized history determines which branch of the rule applies. Nor is it the same as a trial-level adaptation, such as changing the protocol after an interim review. A trial-level change may cause different participants to be governed by different patient-level regimes.
Defining the Regime-Specific Estimand
Let \(Y^d\) denote the outcome that would be observed for a patient under the complete dynamic regime \(d\), under the specified outcome and intercurrent-event strategy. For a fixed target population \(P^*\), the mean outcome under that regime is
A comparison of two regimes \(d_A\) and \(d_B\), for a mean outcome, can be expressed as
The target population, outcome variable, summary measure, and strategy for intercurrent events must be held clear when asking whether a rule revision changed the estimand. This tutorial focuses on the treatment condition: the regime itself. The formula does not assert that the mean is identifiable from any particular trial data. Estimation also requires suitable design and assumptions.
For a sequential rule to be evaluated from trial data, treatment assignment at each decision time must be randomized or otherwise conditionally exchangeable given the measured history, and each treatment required by the rule must have positive probability for histories in the target population. The observed history must be measured before the treatment decision it informs. In addition, the observed outcome under a participant’s assigned strategy must correspond to the outcome under that strategy. These are identification considerations; they do not change what the estimand is, but their failure can prevent reliable estimation of it.
When Does Revising the Rule Change the Estimand?
Two rules define the same treatment condition for a target population if they prescribe the same treatment at every decision point for histories that can occur in that population, including the resulting future branches. A revision that changes only wording or changes a branch that cannot occur in the target population may leave the regime unchanged for that target. A revision that changes treatment for a history with positive probability generally defines a different regime.
The implication for a mean outcome follows by separating the population into histories where the rules agree and histories where they differ. Suppose the rules first diverge for a subgroup \(G\), and they agree for everyone outside \(G\). If the subgroup has population proportion \(q\), and the rules’ mean outcomes in that subgroup are \(\mu_A\) and \(\mu_B\), then
provided the regimes yield the same mean outcome outside \(G\). Thus a changed branch alters the regime, while the resulting numerical difference in the estimand depends on how often that branch occurs and how outcomes differ under the two actions. It is possible for two distinct regimes to have the same mean outcome by coincidence; that does not make their treatment conditions identical.
Worked Example: Revising a Response Threshold
Hypothetical data. A trial’s fixed target population has 100 patients. At week 4, 30 are classified as poor responders. Regime \(d_A\) switches every poor responder to a second treatment; regime \(d_B\) switches only the 12 patients with very poor response and continues the initial treatment for the other 18 poor responders. The rules agree for the remaining 70 patients. Suppose the mean outcome score in the 18 intermediate responders is 68 under \(d_A\) and 74 under \(d_B\), and the other patients have the same mean outcome under both regimes.
The rules differ for \(18/100=0.18\) of the target population. Their contrast in mean outcome is therefore
Under these hypothetical values, the mean outcome under \(d_A\) is 1.08 points lower than under \(d_B\). Revising the threshold changes the treatment condition because it changes the treatment for a group present in the target population. The rule change alone does not establish which strategy is preferable; that conclusion depends on the outcome definition, uncertainty, and clinical interpretation.
Worked Examples of Preserving or Changing the Treatment Condition
Worked Example: A Revision Outside the Target Population’s Histories
Hypothetical data. Consider a fixed target population of 200 patients in which none has a contraindication recorded at the treatment decision. Rule \(d_A\) assigns treatment X unless a contraindication is present, in which case it assigns treatment Y. A revised rule \(d_B\) assigns treatment X unless a contraindication is present, in which case it assigns treatment Z. In this target population, the observed history has no contraindication for all 200 patients, and the mean outcome under treatment X is 50 points.
For every history in this target population, both rules assign X. Thus
For this target population, the rules induce the same treatment condition on the histories that occur. They would not be equivalent for a target population that includes patients with contraindications. The conclusion is therefore about the specified target population, not a universal claim that the rules are identical.
Worked Example: Pooling Participants From Two Rule Versions
Hypothetical data. Suppose 60 patients in an early trial period follow rule \(d_0\), with a mean outcome of 70 points, and 40 patients in a later period follow revised rule \(d_1\), with a mean outcome of 74 points. For this calculation, assume both periods represent the same target population and that the reported means estimate the respective regime-specific population means.
The pooled mean across all 100 patients is
Under the stated assumptions, this pooled summary corresponds to a policy that assigns \(d_0\) with probability 0.60 and \(d_1\) with probability 0.40, independently of patient characteristics in the target population. It is not, in general, the mean outcome under \(d_0\) alone or under \(d_1\) alone. If period and patient characteristics differ, or if the rule version is tied to calendar time, the pooled mean need not represent that simple randomized mixture. The trial should then state the actual treatment condition and target explicitly rather than treating a pooled summary as if it were the effect of one unchanged rule.
Distinguishing Patient-Level Rules From Trial-Level Adaptation
The following distinction helps identify what treatment condition a trial actually compares.
| Feature | Patient-level dynamic regime | Trial-level adaptation |
|---|---|---|
| What determines treatment? | The individual patient’s observed history at a decision time | A trial decision, such as a protocol revision after an interim review |
| Example | Switch if the patient’s week-4 response is below a defined threshold | After interim review, use a revised threshold for later enrollees |
| Potential estimand issue | A different mapping from history to treatment defines a different regime on affected histories | Participants may be governed by different regimes; pooling can target a mixture or another explicitly defined policy |
A protocol revision need not change the estimand if it leaves the mapping unchanged for all histories in the target population. Conversely, a change described as “operational” can change the estimand if it changes which treatment patients receive under some relevant history. The key question is not whether the protocol document changed, but whether the treatment rule defining the condition changed.
Diagnostics and Common Pitfalls
Before concluding that an adaptation preserves a treatment condition, compare the old and new rules branch by branch. Specify the timing of each decision, the permitted information at that time, the response definition and threshold, and the treatment assigned under each possible history. Then ask whether each changed branch occurs in the fixed target population and whether it changes the treatment prescribed. This is a diagnostic of the treatment condition, not a substitute for evaluating whether the trial data can estimate its outcome.
- Unspecified history: A rule such as “switch for inadequate response” is not reproducible until response, measurement time, and threshold are defined.
- Unrecorded decision information: If the rule uses a history variable that is not measured, it may be impossible to verify who followed which branch or to support estimation under the regime.
- Unsupported treatment choices: If no patients with a relevant history could receive the treatment prescribed by the rule, the required positivity condition fails for that history.
- Calendar-time confounding: If a revised rule is used only later in the trial, changes over time in patient mix, care, or outcome measurement can complicate comparisons of rule versions.
A common mistake is to reason, “The trial still studies the same drug and measures the same endpoint, so the estimand is unchanged.” That reasoning omits the treatment condition. The correction is to compare the complete history-to-treatment mappings. If the revised rule changes treatment for a history represented in the target population, the treatment condition has changed, even when the drug names and endpoint are unchanged.
A second pitfall is to conclude that two distinct rules are the same regime because their estimated mean outcomes happen to be equal. Equality of means is not equality of interventions: the rules may assign different treatments to some patients while producing the same average outcome. Conversely, a changed branch can define a different regime even when its population frequency is small.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A rule assigns treatment A to patients with a week-2 response of at least 5 units and treatment B otherwise. Name the history-dependent feature that makes this a dynamic regime.
- In a target population, 25% of patients follow a branch altered by a rule revision. Their mean outcome is 3 points higher under the revised rule, while all other branches have equal mean outcomes. What is the revised-minus-original difference in population mean?
- A rule is changed for a history that cannot occur in the specified target population. Does this necessarily change the treatment condition for that target population? Explain.
- Forty percent of a target population follows rule \(d_0\), with mean outcome 60 points; the remaining 60% follows rule \(d_1\), with mean outcome 65 points. Under the same-population and mixture assumptions used above, calculate the pooled mean and identify its treatment-policy interpretation.
Solutions. (1) Treatment depends on the patient’s observed week-2 response, measured before the decision; the threshold and the action on each side define the rule. (2) The difference is \(0.25(3)=0.75\) points in favor of the revised rule. (3) No. If the changed history cannot occur in the target population and the rules agree on all histories that can occur, they induce the same treatment condition for that target population; the conclusion may differ for another population. (4) The pooled mean is \(0.40(60)+0.60(65)=24+39=63\) points. Under the stated assumptions, it is the mean for a policy assigning \(d_0\) with probability 0.40 and \(d_1\) with probability 0.60, not the mean under either rule alone.