Two Goals That Can Pull in Different Directions
A trial allocation rule can affect both the outcomes experienced by participants in the trial and the amount of information the trial provides about treatment differences. If evidence suggests that one treatment is more effective, assigning more participants to it may increase the expected number who respond. But an unequal split can also make the comparison less precise than an equal split with the same total sample size. These are related consequences of allocation, not interchangeable measures of success.
The introductory material on the value of information in treatment assignment explains why treatment assignment can serve both participants and future decision-making. Here we make one part of that tension concrete using a binary outcome: each participant either responds or does not respond. The calculations isolate the allocation trade-off; they do not by themselves establish that a response-adaptive design improves outcomes overall.
Assumptions for the Basic Calculations
For the examples, suppose every participant has a binary outcome observed completely, and each treatment has a stable response probability: participants assigned to A respond independently with probability \(p_A\), and participants assigned to B respond independently with probability \(p_B\). Assume also that the two treatment groups’ outcomes are independent. These assumptions give \(X_A\sim\operatorname{Binomial}(n_A,p_A)\) and \(X_B\sim\operatorname{Binomial}(n_B,p_B)\), where \(X_A\) and \(X_B\) are the response counts.
The expected number of responses follows by adding the expected response counts in the two groups. For a fixed allocation, the estimated treatment difference is \(X_B/n_B-X_A/n_A\). Because a sample proportion based on \(n\) independent binary outcomes has variance \(p(1-p)/n\), independence gives the variance and standard error of the difference:
In practice, \(p_A\) and \(p_B\) are not known. The variance formula therefore describes the sampling precision under specified response probabilities; an analysis would estimate its standard error using observed data and an appropriate method. In the examples, the probabilities are supplied to make the allocation consequences transparent. They are hypothetical values, not results from an actual study.
Worked Allocation Comparisons
Worked Example: More Assignments to the More Effective Treatment
Suppose a hypothetical trial has 60 participants and two binary-outcome treatments. For illustration, take the response probabilities to be \(p_A=0.40\) and \(p_B=0.60\). Compare equal allocation, with 30 participants per arm, to an unequal allocation with 20 assigned to A and 40 to B.
Expected responses with equal allocation. Substituting the group sizes and response probabilities gives
Expected responses with unequal allocation.
Under these stipulated probabilities, the unequal allocation increases the expected total by 2 responses among the 60 trial participants. This is an expectation, not a promise that the realized trial will have exactly 32 responses.
Precision of the comparison. With equal allocation, the variance of the difference in sample response proportions is
With 20 assigned to A and 40 to B, it is
Thus, in this example, allocating more participants to B raises expected responses but also raises the standard error of the estimated response-proportion difference. The units of the standard error are proportion points; \(0.1342\) corresponds to about 13.4 percentage points.
The calculation separates two questions. “How many responses do we expect among trial participants?” depends on the allocation-weighted average of the treatment response probabilities. “How precisely can we estimate the difference?” depends on the outcome variability in each arm and the number assigned to each arm. A favorable answer to the first question does not automatically imply a favorable answer to the second.
Worked Example: A Small Apparent Advantage
Suppose interim evidence makes B look only slightly better. For a planning illustration, use \(p_A=0.55\) and \(p_B=0.60\) as provisional response probabilities for 30 future participants. Compare equal allocation (15 per arm) with a 10-to-20 split favoring B.
With equal allocation, the expected number of future responses is
With 10 assigned to A and 20 to B, it is
The projected gain is \(17.5-17.25=0.25\) responses on average. It is a small difference, and the probabilities used are provisional rather than known truths.
Using those same probabilities, the standard error for equal allocation is
For the 10-to-20 allocation, the corresponding calculation is
The projected response gain is modest, while the comparison is less precise under this calculation. This does not tell us how to balance the goals: it makes clear which quantities must be considered together, and how uncertain estimates of response probabilities affect any claim about projected benefit.
Why a Promising Treatment Can Be the Wrong One to Favor
Response-adaptive allocation uses accumulating outcome information to alter assignment probabilities. The preceding examples instead compare allocations under stipulated response probabilities. That distinction matters: in a real trial, the treatment that looks promising at an interim point may not truly have the higher response probability. When the evidence is noisy, increasing assignments to the apparently better arm can direct participants toward the treatment with the lower true response probability.
Worked Example: The Promising Treatment Is Actually Worse
Consider 60 hypothetical participants, with true response probabilities \(p_A=0.60\) and \(p_B=0.40\). Suppose interim data nevertheless make B appear promising, so the remaining allocation favors B. Compare 30 participants per arm with 20 assigned to A and 40 to B.
Under equal allocation, the expected response count is
Under the allocation favoring B, it is
Here the shift lowers the expected number of responses by 2, because B is in fact less effective. The equal-allocation standard error is \(\sqrt{0.24/30+0.24/30}=0.1265\); the unequal-allocation standard error is \(\sqrt{0.24/40+0.24/20}=0.1342\). Both expected outcomes and the comparison’s precision are worse under this particular shift. The example shows why a gain calculated from estimated response rates must not be described as a guaranteed patient benefit.
These calculations assume fixed arm sizes and known response probabilities for the purpose of comparison. In a response-adaptive trial, assignment probabilities and the eventual arm sizes can depend on previous outcomes. Consequently, a fixed-allocation standard error calculation is not a complete analysis of the adaptive design, and projected gains based on interim estimates do not capture all uncertainty in the adaptation process. Design evaluation must consider the allocation rule and its behavior over possible trial outcomes, not just one favorable sequence.
Derivation and Interpretation
The expected response calculation is a direct consequence of indicator variables. Write the response for participant \(i\) as \(Y_i\), where \(Y_i=1\) for a response and \(0\) otherwise. Under a fixed allocation, the total response count is the sum of the individual indicators. For each participant assigned to treatment \(j\), \(E(Y_i)=p_j\). Linearity of expectation therefore gives the sum of the relevant response probabilities, whether or not the indicators are independent:
Independence is needed for the simple variance calculation above: it lets the variances of the two independent sample proportions add without covariance terms. If outcomes are correlated, response probabilities vary across participants, or outcomes are missing, the displayed standard error may not represent the actual uncertainty. The assumptions should be checked against the trial design and data collection before using the formula to support a precision claim.
Comparison, Diagnostics, and Limits
Equal allocation is a useful reference because it treats the two arms symmetrically and, under many common conditions, supports a direct comparison. Unequal fixed allocation can be chosen in advance for reasons such as existing evidence or participant considerations. Response-adaptive allocation differs because assignments change in response to accumulating outcomes. All three approaches can be assessed in terms of expected outcomes and comparative information, but these simple calculations do not establish which allocation is best in every setting.
Several checks help prevent overinterpretation:
- Check the direction of the assumed benefit. Confirm which treatment has the larger response probability in the calculation. A shift toward the lower-probability treatment can reduce expected responses.
- Separate estimates from truths. Interim response proportions are uncertain estimates. A projected gain using them is conditional on those estimates being good approximations.
- Check that the comparison remains estimable. Very few assignments to one arm can make its response proportion highly variable and the treatment comparison weakly informed.
- Check outcome completeness and comparability. Differential missing outcomes or changes in patient mix can undermine calculations based on stable, fully observed binary outcomes.
- Do not treat a fixed-allocation calculation as a full adaptive-design evaluation. When assignment depends on observed outcomes, the entire allocation process and its uncertainty matter.
A common mistake is to reason, “B looks better, so assigning more participants to B must improve the trial without a statistical cost.” The correction is twofold: the apparent advantage may be wrong, and an unequal allocation can reduce precision for the treatment comparison. Conversely, emphasizing precision alone is not a complete ethical argument; the expected experiences of participants also matter. Allocation choices require making both consequences explicit while acknowledging uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- With \(p_A=0.30\), \(p_B=0.50\), and 40 participants, calculate expected responses under equal allocation and under 10 assignments to A and 30 to B.
- For the same probabilities and allocations in Question 1, calculate the standard error of \(\hat{p}_B-\hat{p}_A\) under each allocation.
- Why is a projected increase in expected responses based on interim estimates not a guaranteed patient benefit?
- Name two assumptions behind the simple standard error calculation, and describe one consequence if an assumption fails.
Solutions. (1) Equal allocation gives \(20(0.30)+20(0.50)=6+10=16\) expected responses. The 10-to-30 allocation gives \(10(0.30)+30(0.50)=3+15=18\), an increase of 2 in expectation under the specified probabilities. (2) The equal-allocation standard error is \(\sqrt{0.50(0.50)/20+0.30(0.70)/20}=\sqrt{0.0125+0.0105}=\sqrt{0.0230}=0.1517\). For the unequal allocation it is \(\sqrt{0.50(0.50)/30+0.30(0.70)/10}=\sqrt{0.00833+0.0210}=\sqrt{0.02933}=0.1713\). The unequal allocation has the larger standard error in this example. (3) Interim estimates are uncertain and may identify the wrong treatment as better; realized responses also vary around their expectations. (4) The calculation assumes, among other things, independent binary outcomes and stable response probabilities within each treatment group. Dependence or changing response probabilities can make the displayed variance formula inaccurate, so the stated standard error may misrepresent precision.