Why Define the Target Before Discussing the Allocation Rule?
In an adaptive trial, assignment probabilities may respond to outcomes observed earlier in the trial. That feature can change how many participants receive each treatment and which outcomes are observed in each arm. It does not, by itself, determine what treatment effect the trial is intended to describe. That target must be stated separately.
This distinction matters because an estimand is a feature of a specified population and a set of treatment conditions. An allocation rule is a procedure for assigning treatments. An estimator is a calculation from the data intended to estimate the estimand. These are related, but they are not interchangeable: changing the allocation rule may change the data available for estimation without changing the target defined in advance.
Potential Outcomes and Fixed-Population Targets
Suppose the target consists of \(N\) specific eligible participants, identified before treatment assignment. Let \(a\) denote a treatment arm in a set of \(K\) arms. For participant \(i\), let \(Y_i(a)\) be the outcome that would be observed at the prespecified time point if that participant received arm \(a\). The outcome could be a symptom score, survival status at a specified time, or another well-defined measure. The values \(Y_i(a)\) for all arms are called that participant’s potential outcomes.
Interpreting these potential outcomes requires that each treatment condition be sufficiently well specified to give it a consistent meaning. It also requires that one participant’s outcome under a treatment be defined without depending on other participants’ treatment assignments; this is the no-interference assumption. Consistency connects the potential outcome to the observed outcome: if participant \(i\) is assigned arm \(a\), the observed outcome is \(Y_i(a)\). In a trial, only one of a participant’s potential outcomes is observed.
The treatment effect is an average of individual treatment effects, \(Y_i(a)-Y_i(b)\). It need not be the effect for every participant. Effects may vary across participants and may even have different signs. The average describes the specified population’s mean contrast, not a promise that each participant benefits by that amount.
The population, treatment conditions, outcome, and outcome time point are all part of the estimand’s meaning. A mean at 30 days is not automatically the same estimand as a mean at 90 days. Nor is an average over all eligible participants automatically the same as an average over one subgroup. The target is fixed only after these choices are specified.
Worked Example: A Two-Arm Contrast in a Fixed Cohort
Consider a hypothetical fixed cohort of four participants. Their potential outcomes are measured in symptom-score points, with larger values representing more symptoms.
| Participant | \(Y_i(0)\), control | \(Y_i(1)\), treatment |
|---|---|---|
| 1 | 4 | 5 |
| 2 | 5 | 8 |
| 3 | 7 | 8 |
| 4 | 8 | 10 |
The fixed-population control mean is \( \mu_0^F=(4+5+7+8)/4=24/4=6 \) symptom-score points. The treatment mean is \( \mu_1^F=(5+8+8+10)/4=31/4=7.75 \) points. Therefore, \( \Delta_{10}^F=\mu_1^F-\mu_0^F=7.75-6=1.75 \) points.
This positive contrast means that, for this specified cohort and outcome definition, the treatment potential-outcome mean is 1.75 points higher than the control potential-outcome mean. It does not say that every participant’s score would be higher under treatment.
To see why the observed arm summaries are distinct from the estimand, suppose the adaptive rule assigns participants 1, 2, and 3 to treatment and participant 4 to control. The observed treatment mean is \((5+8+8)/3=7\) points, and the observed control mean is 8 points, giving an observed difference of \(7-8=-1\) point. This realized difference is not the finite-population contrast of 1.75 points. The example makes no claim about the statistical properties of that observed difference; it shows that an observed summary and the target are different quantities.
Why Adaptation Does Not Redefine a Prespecified Estimand
Let \(A_i\) be the arm assigned to participant \(i\), and let \(Y_i^{\mathrm{obs}}\) be the observed outcome. Under consistency, \(Y_i^{\mathrm{obs}}=Y_i(A_i)\). An adaptive design may choose the next participant’s arm using earlier assignments and outcomes. Write the conditional assignment probability as \(P(A_i=a\mid H_i)\), where \(H_i\) denotes the information available when that assignment is made.
The assignment probabilities govern which potential outcomes become observed. They are not terms in the fixed-population definition of \(\mu_a^F\) or \(\Delta_{ab}^F\). The algebraic justification is direct: subtract the two fixed-population means, each defined over the same \(N\) participants.
No assignment probability appears in this identity. Conditional on the specified population and its potential outcomes, the right-hand side is a fixed quantity. Randomness in the adaptive assignment sequence can affect which terms are observed, and therefore the information available to estimate the quantity, but it does not change the quantity’s definition. This is a distinction between the target and the observation process, not a claim that every estimator has desirable properties under every adaptive design.
Worked Example: Three Arm Means Are Not Weighted by Allocation Shares
Suppose a hypothetical fixed population contains five participants, and the outcome is improvement in a clinical score, measured in points. The potential outcomes under arms A, B, and C are:
| Participant | A | B | C |
|---|---|---|---|
| 1 | 2 | 4 | 3 |
| 2 | 3 | 5 | 3 |
| 3 | 4 | 4 | 6 |
| 4 | 5 | 7 | 4 |
| 5 | 6 | 5 | 5 |
The arm-specific means are \( \mu_A^F=(2+3+4+5+6)/5=20/5=4 \) points, \( \mu_B^F=(4+5+4+7+5)/5=25/5=5 \) points, and \( \mu_C^F=(3+3+6+4+5)/5=21/5=4.2 \) points.
Thus, for example, the B-versus-A contrast is \( \Delta_{BA}^F=5-4=1 \) point. Now suppose an adaptive design’s realized assignment shares are 0.2 for A, 0.6 for B, and 0.2 for C. Weighting the three arm means by those shares gives \(0.2(4)+0.6(5)+0.2(4.2)=0.8+3+0.84=4.64\) points. This is an allocation-weighted summary, not the mean under any one arm and not the B-versus-A treatment effect. The arm means remain defined over all five participants under each specified treatment condition.
The Population Is Part of the Estimand
Although assignment probabilities do not enter a fixed-population estimand, changing the target population can change it. This is especially important when treatment effects differ across subgroups. A change in who is targeted is different from a change in which arm is more often assigned within the same target.
Worked Example: Different Target Compositions, Different Average Effects
Suppose a hypothetical population has two equally sized subgroups of five participants each. The outcome is improvement in points, and the treatment-versus-control effect is 4 points for each participant in subgroup H and \(-2\) points for each participant in subgroup L. For the ten-person fixed population, the average effect is \( \{5(4)+5(-2)\}/10=(20-10)/10=1 \) point.
Now consider a different target population whose subgroup composition is 80% H and 20% L, with the same subgroup-specific effects. Its average effect is \(0.8(4)+0.2(-2)=3.2-0.4=2.8\) points. The average differs because the target population differs, not because the assignment probabilities changed. Assigning more treatment to H within the original ten-person population would not, by itself, change that original population’s 1-point average effect.
Choosing Among Related Quantities
Several quantities can be reasonable to report, but they answer different questions. State which one is intended rather than treating them as interchangeable.
| Quantity | What it describes | Role of assignment shares |
|---|---|---|
| Finite-population arm mean, \( \mu_a^F \) | Mean potential outcome if the fixed target population received arm \(a\) | Not part of its definition |
| Finite-population contrast, \( \Delta_{ab}^F \) | Difference between two arm-specific means in that population | Not part of its definition |
| Observed mean in an assigned arm | Mean of the outcomes actually observed among participants assigned to that arm | Can affect which participants contribute observed outcomes |
| Allocation-weighted mean of arm means | A mixture of arm means using stated weights | The weights define the mixture, but it is not an arm-specific mean |
An allocation-weighted mean may be relevant when the question concerns a particular mixture of treatments. In that case, the weights should be stated as part of the target. They should not be silently substituted for the equal-participant average treatment effect in the fixed population.
Assumptions, Diagnostics, and Limitations
The potential-outcome definitions are meaningful only when the target and treatment conditions are clear. Check that the population is specified, the outcome and time point are fixed, and treatment versions are sufficiently well defined. If one participant’s outcome can be affected by another participant’s assignment, the no-interference formulation needs reconsideration or extension. If outcomes are undefined for some participants under a treatment condition, the estimand needs an explicit rule for that situation rather than an unspoken assumption.
A useful diagnostic is to write the estimand before looking at the realized arm counts. If its definition changes when those counts change, ask whether the target itself was intentionally defined using allocation shares or whether the design has been confused with the target. Also check whether subgroup composition is part of the population definition: an average over a fixed cohort and an average over a differently composed target are not interchangeable.
A finite-population estimand describes the specified participants; it does not automatically describe a broader population from which they might have been drawn. A broader-population target requires specifying that population or distribution. In addition, defining a causal contrast does not ensure that the trial data identify it precisely or that a chosen estimator is unbiased. Those are questions about the observation and estimation procedure. The estimand tells us what quantity those procedures are meant to address.
Check Your Understanding
Work through each question, then compare your reasoning with the solutions.
- A fixed cohort of four has potential outcomes under control of 2, 4, 6, and 8, and under treatment of 3, 5, 9, and 9. Calculate both arm means and the treatment-minus-control average effect.
- In a three-arm trial, an adaptive design assigns 10%, 70%, and 20% of participants to arms A, B, and C. Does that allocation pattern change the definition of the fixed-population mean under B? Explain.
- Two subgroups have treatment effects of 3 and \(-1\) points. The target is 25% from the first subgroup and 75% from the second. Calculate the target average effect.
- Explain why the observed mean among participants assigned to treatment is not, by definition, the treatment potential-outcome mean for the whole fixed population.
Solutions. 1. The control mean is \( (2+4+6+8)/4=20/4=5 \). The treatment mean is \( (3+5+9+9)/4=26/4=6.5 \). The treatment-minus-control effect is \(6.5-5=1.5\) outcome units. 2. No. The fixed-population mean under B is the average of \(Y_i(B)\) over every participant in the specified population. The assignment shares determine the allocation pattern, not the terms or denominator in that definition. 3. The average is \(0.25(3)+0.75(-1)=0.75-0.75=0\) points. 4. The observed treatment mean averages outcomes only for those assigned treatment, while the fixed-population treatment mean averages the treatment potential outcome over the entire target population. These are different definitions; the observed mean does not include unobserved treatment potential outcomes for participants assigned elsewhere.