When Unequal Allocation Does Not Cost Much Power

Power and Patient Outcomes Are Different Design Goals

Unequal allocation can direct more participants to a treatment expected to provide greater benefit, while leaving fewer participants in the other group to support a precise treatment comparison. The practical question is not simply whether imbalance reduces power: it does, under the usual equal-variance model and fixed total sample size. The question is whether the reduction is small enough to be acceptable in light of the corresponding expected patient-outcome gain.

The examples here concern a continuous outcome for which larger values are better. They use a two-sided test of a difference in group means. This setting makes the trade-off transparent; it does not imply that every response-adaptive design has the same power as a fixed allocation with the same final group sizes. In a response-adaptive trial, the allocation process and its uncertainty may matter as well.

Model and assumptions: Let group A and group B have independent outcomes with normally distributed responses, means \(\mu_A\) and \(\mu_B\), and a common known standard deviation \(\sigma\). The total sample size \(N=n_A+n_B\) and group sizes are treated as fixed. The test is two-sided, with \(H_0:\mu_B-\mu_A=0\), \(H_a:\mu_B-\mu_A\ne0\), and significance level \(\alpha=0.05\). Under an alternative with \(\mu_B-\mu_A=\delta>0\), the difference in sample means has standard error \(SE=\sigma\sqrt{1/n_A+1/n_B}\). These calculations assume independent observations and no loss of information from missing outcomes or other design features.

How Imbalance Changes the Standard Error and Power

For this model, the standardized difference in sample means has a normal distribution with variance 1 and mean \(\lambda=\delta/SE\) under the stated alternative. Write \(\Phi\) for the standard normal cumulative distribution function and \(z_{1-\alpha/2}\) for its \(1-\alpha/2\) quantile. The test rejects when the absolute standardized statistic exceeds \(z_{1-\alpha/2}\). Its power is the probability of rejection when the stated alternative is true:

\text{Power}=\Phi(\lambda-z_{1-\alpha/2})+\Phi(-\lambda-z_{1-\alpha/2}).

At \(\alpha=0.05\), \(z_{1-\alpha/2}\) is approximately 1.960. Power depends on the standardized effect \(\lambda\), not on the allocation ratio by itself. Consequently, the same imbalance can have a small absolute power cost in a study with a clear signal, but a more consequential cost when power is already near a decision threshold.

Let \(r=n_B/n_A\), so that \(r>1\) means more participants are assigned to B. With \(N\) fixed, \(n_A=N/(1+r)\) and \(n_B=Nr/(1+r)\). Substitution into the variance formula gives a useful measure of the allocation penalty.

SE^2=\sigma^2\left(\frac{1}{n_A}+\frac{1}{n_B}\right) =\frac{\sigma^2}{N}\left(r+2+\frac{1}{r}\right).

For equal allocation, \(r=1\), and \(SE^2=4\sigma^2/N\). Dividing the unequal-allocation variance by this equal-allocation variance gives the relative variance inflation \((r+1)^2/(4r)\). Since \(r+1/r\ge2\), this factor is at least 1 and is exactly 1 only for equal allocation. The standard error is inflated by the square root of this factor. The relation is symmetric: ratios \(r\) and \(1/r\) produce the same variance penalty when the total sample size is fixed.

Key interpretation: Variance inflation is not the same as a percentage-point loss in power. To find the power change, recalculate \(SE\), then \(\lambda\), then the rejection probability. The power difference depends on the effect size, sample size, significance level, and allocation ratio together.

Worked Comparisons

Worked Example: A 3:2 Allocation With a Small Power Cost

Suppose a hypothetical trial plans to compare two treatments on a continuous outcome measured in points. Assume independent normal outcomes, a common known standard deviation of 12 points, and a true mean difference of 4 points in favor of B. There are 240 participants. Compare equal allocation with assigning 144 to B and 96 to A. These assumptions match the model above; in particular, the group sizes are fixed for this calculation.

With 120 participants per group, the standard error is \(12\sqrt{1/120+1/120}=1.5492\) points, and \(\lambda=4/1.5492=2.5820\). The power is \(\Phi(2.5820-1.960)+\Phi(-2.5820-1.960)\), approximately 0.7330.

With 144 participants in B and 96 in A, \(SE=12\sqrt{1/144+1/96}=1.5811\) points, and \(\lambda=4/1.5811=2.5298\). The power is \(\Phi(2.5298-1.960)+\Phi(-2.5298-1.960)\), approximately 0.7156. The decrease is \(0.7330-0.7156=0.0174\), or 1.74 percentage points.

Now compare expected outcomes received by participants. If the hypothetical means are 50 points for A and 54 for B, equal allocation has an expected average of \((50+54)/2=52\) points per participant. The unequal allocation has expected average \((96(50)+144(54))/240=52.4\) points. Thus the planned allocation increases the expected average outcome by 0.4 points per participant, or 96 points summed across 240 participants. This is an expectation under the assumed means, not a guaranteed improvement for the enrolled group.

Worked Example: The Same Moderate Imbalance Under a Stronger Signal

Consider a separate hypothetical trial with 200 participants, common known standard deviation 10 outcome units, and a true difference of 5 units favoring B. Outcomes are independent and normally distributed, the variance is common, and the group sizes are fixed. Compare 100:100 with 120 in B and 80 in A.

Under equal allocation, \(SE=10\sqrt{1/100+1/100}=1.4142\), so \(\lambda=5/1.4142=3.5355\). The two-sided power is \(\Phi(3.5355-1.960)+\Phi(-3.5355-1.960)\), approximately 0.9424.

Under 120:80 allocation, \(SE=10\sqrt{1/120+1/80}=1.4434\), so \(\lambda=5/1.4434=3.4641\). Power is \(\Phi(3.4641-1.960)+\Phi(-3.4641-1.960)\), approximately 0.9337. The decrease is about 0.0087, or 0.87 percentage points.

If the hypothetical means are 40 units for A and 45 for B, equal allocation gives an expected average of 42.5 units. Assigning 60% to B gives \(0.4(40)+0.6(45)=43.0\) units, a gain of 0.5 units per participant on average, or 100 units summed over the trial. This example illustrates why the same 3:2 ratio can have a relatively small absolute power cost when the assumed signal is strong. It does not establish that such a gain is worth the inferential cost in any particular trial.

Worked Example: A More Pronounced Imbalance Has a Larger Cost

For scale, suppose a hypothetical study again has 240 participants, a common known standard deviation of 12 points, and a true 4-point advantage for B. Compare 120:120 with 160 in B and 80 in A. As before, outcomes are assumed independent and normal with equal variances, and the group sizes are fixed.

The equal-allocation calculation gives power 0.7330. With 160:80 allocation, \(SE=12\sqrt{1/160+1/80}=1.6432\) points and \(\lambda=4/1.6432=2.4343\). Thus power is \(\Phi(2.4343-1.960)+\Phi(-2.4343-1.960)\), approximately 0.6823. The decrease is 0.0507, or 5.07 percentage points.

If the assumed means remain 50 points for A and 54 for B, the expected average under 160:80 is \((80(50)+160(54))/240=52.667\) points. This is 0.667 points per participant above the equal-allocation expectation of 52 points. The larger patient-outcome gain comes with a visibly larger power decrease. Whether that trade-off is acceptable depends on the trial’s inferential objective and on how credible the assumed treatment difference is.

How to Evaluate the Trade-Off in Practice

A disciplined comparison keeps the inferential and patient-centered quantities separate before judging them together. First calculate the standard error and power under a common set of assumptions. Then calculate the expected outcome under each allocation using the assumed treatment means. Report the power change in percentage points and the expected outcome change in its natural units. A statement such as “power falls by 2%” is ambiguous unless it specifies whether this means a relative percent change or a percentage-point change.

Expected patient outcomes are particularly sensitive to the assumed ordering and size of the treatment means. If B is expected to be better and receives more participants, the expected average outcome rises under that assumed model. If the assumed advantage is wrong, or reverses, assigning more participants to B may lower rather than raise the expected average. The calculation is therefore a design comparison conditional on assumptions, not evidence that B is superior.

The equal-variance formula is most useful as a benchmark when outcome variability is reasonably similar across groups. If the variances differ, the appropriate variance of the difference is \(\sigma_A^2/n_A+\sigma_B^2/n_B\), not \(\sigma^2(1/n_A+1/n_B)\); the allocation that best supports precision can then differ from equal allocation. Also check that the planned total sample size and the group counts used in the calculation are feasible, and that the outcome scale and direction are interpreted consistently.

In a response-adaptive trial, final group sizes are not generally fixed in advance: they depend on the allocation rule and the evolving trial history. A fixed-count calculation can still illustrate a conditional precision trade-off, but it is not by itself a complete power analysis for an adaptive design. The fixed-allocation benchmark and matched-design comparisons introduced earlier help clarify what is being held constant. A fuller assessment must account for the actual allocation mechanism and the intended inferential procedure.

Common Mistake and Limitations

Wrong reasoning: “A 3:2 ratio inflates variance by a known amount, so power must fall by that same percentage.” Correction: variance inflation changes the standard error, which changes the standardized effect; power must then be recomputed from the rejection probability.
Wrong reasoning: “A higher expected average outcome proves that the adaptive allocation benefits patients.” Correction: the calculation is conditional on the assumed treatment means and the rule assigning more participants to B. It neither proves which treatment is better nor guarantees a realized benefit.
Wrong reasoning: “The examples show that unequal allocation is harmless.” Correction: they show that some moderate imbalances can have small absolute power costs under particular effect sizes and sample sizes. Other settings, including weaker signals or more extreme imbalance, can have larger costs.

Finally, these calculations use a common known standard deviation and a normal test to isolate the allocation effect. Real analyses may estimate variances, use different outcome models, face missing data, or use an allocation procedure that makes the realized counts random. Each can alter the comparison. The calculations are a transparent first check, not a substitute for evaluating the planned trial under its actual design and analysis.

Check Your Understanding

Answer each question using the assumptions and distinctions developed above. Brief solutions follow.

For a fixed total sample size and common variance, why does unequal allocation increase the variance of the difference in sample means?
In the first worked example, what is the absolute power decrease from 120:120 to 144:96 allocation?
If B’s assumed mean is higher than A’s, why can assigning more participants to B raise expected average outcomes without proving B is superior?
Why is a fixed-count power calculation not automatically a full power analysis for a response-adaptive design?

Solutions. (1) The variance is \(\sigma^2(1/n_A+1/n_B)\); at fixed \(N\), it is minimized when the group sizes are equal, as shown by \((r+1)^2/(4r)\ge1\). (2) It is \(0.7330-0.7156=0.0174\), or 1.74 percentage points. (3) The expected average is a weighted average of the assumed means; shifting weight toward the assumed higher mean raises that calculation, but the assumed ordering may be wrong and the result is not guaranteed for a realized trial. (4) Adaptive group sizes depend on the allocation process and trial history, so the fixed-count calculation does not account for their randomness or the analysis procedure.