Why Treatment-Effect Differences Matter
An allocation rule determines which treatment a patient is assigned, and the chance of assignment may depend on information known before randomization. If treatment effects are the same across patients, changing allocation between patient groups does not by itself create more expected treatment benefit when the overall fraction assigned to treatment is held fixed. If effects differ across groups, however, who receives treatment can matter: a rule that assigns treatment more often to a group with a larger treatment effect can increase the trial population’s expected outcome.
This comparison is about expected outcome under an allocation rule, not a guarantee that any particular patient will benefit. A patient’s outcomes under both treatments cannot be observed, so an individual causal effect is generally not known. Subgroup treatment effects are group averages, and even a group with a positive average effect can include patients who do not benefit.
Let \(\pi_g=P(G=g)\) be the proportion of the population in subgroup \(g\), where \(\pi_1+\pi_2=1\). Write \(\mu_{0g}=E[Y(0)\mid G=g]\). Under an allocation rule that assigns treatment with probability \(q_g\) in subgroup \(g\), the expected outcome in that subgroup is
Consequently, the population-weighted expected outcome is
This expression separates two contributions. The baseline term \(\sum_g\pi_g\mu_{0g}\) does not depend on allocation. The allocation-dependent contribution, \(\sum_g\pi_gq_g\tau_g\), depends on both the treatment probabilities and the subgroup-specific effects. When comparing rules for the same population, the baseline term cancels.
Assumptions for the Comparison
These calculations require a clear target population and a common outcome scale. The subgroup proportions \(\pi_1,\pi_2\) must describe that population; they may be known by design or stipulated for a hypothetical calculation. Subgroup membership must be available before treatment assignment, and the allocation rule must use it as stated. The effects and baseline means must refer to the same outcome, follow-up time, and patient population.
The potential-outcome setup also assumes that each patient’s observed outcome is the outcome under the treatment actually assigned, and that one patient’s assignment does not change another patient’s outcome. Randomization within each subgroup makes treatment assignment independent of potential outcomes within that subgroup, in expectation. For the examples below, subgroup means and effects are supplied as known hypothetical values. In practice, they are uncertain estimates; treating estimates as known can make a proposed rule appear more beneficial than it is.
A Two-Subgroup Comparison
Suppose the rule must assign a fixed overall proportion \(\bar q\) of patients to treatment. An allocation rule is feasible if its subgroup probabilities satisfy
A subgroup-blind rule uses the same probability in both groups, \(q_1=q_2=\bar q\). A subgroup-specific rule can use different probabilities while meeting the same overall treatment fraction. Holding \(\bar q\) fixed makes the comparison informative about how treatment is distributed across groups, rather than simply how many patients receive treatment.
Worked Example: More Treatment for the Higher-Effect Subgroup
Consider a hypothetical population with equal subgroup proportions, \(\pi_1=\pi_2=0.50\). On a clinical score where larger is better, suppose the known effects are \(\tau_1=8\) points and \(\tau_2=2\) points. Compare fixed 1:1 allocation, \((q_1,q_2)=(0.50,0.50)\), with a subgroup-specific rule, \((q_1,q_2)=(0.80,0.20)\). Both assign half the population to treatment overall.
The treatment allocation contribution under fixed allocation is
Under subgroup-specific allocation it is
The allocation-dependent expected outcome is therefore higher by \(3.40-2.50=0.90\) points per patient, relative to the common baseline contribution, under the subgroup-specific rule. This is an expected population-level difference under the stipulated effects and proportions, not a claim that each patient gains 0.90 points or that every patient in subgroup 1 benefits.
Why Alignment Changes Expected Benefit
The key comparison can be derived directly. Under the fixed treatment-fraction constraint, the allocation-dependent contribution for a subgroup-specific rule is \(\pi_1q_1\tau_1+\pi_2q_2\tau_2\). For the subgroup-blind rule it is \(\bar q(\pi_1\tau_1+\pi_2\tau_2)\). Since \(\bar q=\pi_1q_1+\pi_2q_2\), subtracting the subgroup-blind contribution and collecting terms gives
Thus, if subgroup 1 has a larger effect than subgroup 2, then assigning treatment more often in subgroup 1 (\(q_1>q_2\)) makes \(\Delta\) positive. If the effects are equal, the difference is zero for any feasible allocation with the same overall treatment fraction. If allocation is tilted toward the subgroup with the smaller effect, the difference is negative. This is a comparison of expected outcomes under specified effects; it does not establish that those effects are known accurately in an actual trial.
Worked Example: Unequal Subgroup Proportions
Suppose subgroup 1 makes up 30% of a hypothetical population and subgroup 2 makes up 70%. Their treatment effects are \(\tau_1=10\) and \(\tau_2=2\) points. Compare a subgroup-blind treatment probability of \(\bar q=0.50\) with the feasible subgroup rule \(q_1=0.80\), \(q_2=0.3714\). The latter satisfies the treatment-fraction constraint because
Under subgroup-blind allocation, the allocation-dependent contribution is
Under the subgroup-specific rule it is
The difference is about \(0.720\) points per patient. The derivation gives the same result:
The unequal subgroup sizes matter: the allocation probabilities must be weighted by the subgroup proportions to check the overall treatment fraction and compute the population expectation.
How to Compare Allocation Rules
A useful comparison proceeds by fixing the target population and the overall treatment fraction, then calculating the treatment contribution under each rule. The following steps make the weighting and interpretation explicit.
Record the subgroup proportions \(\pi_1,\pi_2\), ensuring they sum to one and represent the population for which benefit is being described.
For each subgroup, state the treatment effect \(\tau_g\) and treatment probability \(q_g\). Check that each probability lies between zero and one.
Compute \(\pi_1q_1+\pi_2q_2\). If the comparison is intended to hold treatment exposure fixed, this quantity must equal the same \(\bar q\) for every rule.
Compute \(\sum_g\pi_gq_g\tau_g\) for each rule and subtract. Report the difference in outcome units per patient, relative to the common baseline contribution.
Worked Example: A Misaligned Allocation Rule
Return to equal subgroup proportions, with effects \(\tau_1=8\) and \(\tau_2=2\) points. A proposed rule assigns treatment with probabilities \(q_1=0.20\) and \(q_2=0.80\). It still assigns half the population to treatment, because \(0.50(0.20)+0.50(0.80)=0.50\), but it allocates treatment more often to the lower-effect subgroup.
The subgroup-specific allocation contribution is
Fixed 1:1 allocation gives \(2.50\) points, as calculated earlier, so the proposed rule is lower by \(1.60-2.50=-0.90\) points per patient relative to the common baseline contribution. The sign also follows from the derived expression:
The calculation is conditional on the stipulated effects being correct. It does not mean that treatment is harmful to every patient in subgroup 2; its average effect is positive in this example, but smaller than subgroup 1’s.
Diagnostics, Alternatives, and Limitations
Before interpreting a claimed benefit, check the inputs and the comparison itself. Frequent warning signs include subgroup proportions that do not describe the target population, allocation probabilities that do not satisfy the stated overall treatment fraction, effects reported on different outcome scales, or subgroup definitions that use information observed after treatment assignment. If the effects are assumed rather than supported by credible evidence, the arithmetic may be correct while the conclusion is fragile.
| Approach | What it does | Key implication |
|---|---|---|
| Fixed, subgroup-blind randomization | Uses the same treatment probability across subgroups. | Provides a transparent reference allocation and does not depend on effect differences between subgroups. |
| Subgroup-specific allocation | Uses different probabilities based on subgroup membership. | Can increase expected outcome at a fixed treatment fraction when allocation is aligned with genuine effect differences. |
| Change the overall treatment fraction | Assigns more or fewer patients to treatment overall. | Changes treatment exposure as well as its distribution, so it is not the same comparison as reallocating a fixed treatment budget. |
The calculation is an expectation over patients, not a prediction of each patient’s outcome. It also assumes the supplied effects do not change as allocation changes. In a real trial, subgroup effects are usually estimated with uncertainty, and subgroup differences can be overstated by chance, especially when many subgroups are examined. A rule based on unstable estimates may direct treatment toward the apparently favorable group even when the true ordering is different.
Finally, expected outcome is only one consideration in choosing an allocation rule. The calculation alone does not account for precision of treatment-effect estimates, participant preferences, safety constraints, or whether an outcome scale captures the benefits that matter to patients. Those considerations require additional evidence and decision criteria.
Common Mistake
- Wrong reasoning: “Subgroup 1 has the larger average effect, so every patient in subgroup 1 will benefit more from treatment.”
- Correction: The calculation compares subgroup-average potential outcomes. It supports a statement about expected outcome across the specified population, under the assumed effects and allocation probabilities. It does not reveal both potential outcomes for an individual patient or guarantee individual benefit.
- Another check: A rule that favors the higher-effect subgroup does not automatically improve expected outcome if it also changes the overall treatment fraction. Separate the effect of redistribution from the effect of assigning more patients to treatment overall.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- Two equally common subgroups have effects of 6 and 6 points. With an overall treatment fraction of 0.50, can reallocating treatment between these subgroups change the expected outcome contribution? Why?
- Subgroup 1 makes up 40% of a population and subgroup 2 makes up 60%. A rule uses \(q_1=0.70\) and \(q_2=0.30\). What overall fraction receives treatment?
- For equally common subgroups with effects \(\tau_1=5\) and \(\tau_2=1\), compare fixed probabilities of 0.50 with \(q_1=0.75\), \(q_2=0.25\). What is the change in expected outcome contribution?
- In one sentence, explain why a positive subgroup-average treatment effect does not imply that every member of that subgroup benefits.
Solutions. (1) No. With equal effects, the derived difference is \(\pi_1\pi_2(q_1-q_2)(\tau_1-\tau_2)=0\), provided the overall treatment fraction is held fixed. (2) The overall fraction is \(0.40(0.70)+0.60(0.30)=0.28+0.18=0.46\), or 46%. (3) The fixed-allocation contribution is \(0.50(0.50)(5)+0.50(0.50)(1)=1.50\) points. The subgroup-specific contribution is \(0.50(0.75)(5)+0.50(0.25)(1)=1.875+0.125=2.00\) points, an increase of \(0.50\) points per patient. (4) A subgroup-average effect summarizes expected outcomes across members; individual treatment and control potential outcomes can differ from that average, and both cannot generally be observed for one patient.