Why Allocation Balance Affects Precision
When a trial compares two treatments using a binary outcome, the estimated treatment contrast is often the difference between the two sample proportions. The number assigned to each treatment affects how precisely that difference can be estimated. With the total number of participants held fixed, moving participants between arms can increase or decrease variance; the direction depends on outcome variability in each arm.
Here, “information” about the contrast means its precision, summarized by the reciprocal of its variance: lower variance means greater precision. This is a useful, specific measure, not a complete account of every kind of information or every ethical consideration involved in allocation.
Deriving the Variance
For a Bernoulli outcome with success probability \(p\), the variance of one observation is \(p(1-p)\). The sample proportion is the average of \(n\) independent Bernoulli observations, so its variance is \(p(1-p)/n\). Applying this result in each arm gives
Because the two sample proportions are independent under the stated model, the variance of their difference is the sum of their variances. Subtracting one random quantity from another does not change the fact that the independent variances add.
Write \(v_A=p_A(1-p_A)\) and \(v_B=p_B(1-p_B)\) for the two arm-specific outcome variances. Then the contrast variance is \(v_A/n_A+v_B/n_B\). The formula makes two points clear: adding participants to an arm reduces that arm’s contribution, while the same number of participants can contribute differently in the two arms if \(v_A\) and \(v_B\) differ.
In practice, the probabilities are usually unknown. A common estimated standard error substitutes the sample proportions for them:
This estimated standard error describes sampling uncertainty under the independent-binomial model. It is not the same as knowing the probabilities and calculating the model-based variance in advance.
Worked Example: Equal Variability and a 70/30 Allocation
Suppose a hypothetical trial plans to observe 100 participants, and planning assumptions set both arm success probabilities to 0.40. Compare a 50/50 allocation with a 70/30 allocation using the model-based variance.
The assumptions give \(v_A=v_B=0.40(1-0.40)=0.24\). With 50 participants in each arm,
The corresponding standard deviation is \(\sqrt{0.0096}\approx 0.0980\), in proportion units. With 70 participants in arm A and 30 in arm B,
The standard deviation is \(\sqrt{0.01143}\approx 0.1069\). Under these equal-variability assumptions, the 70/30 allocation has about \(0.01143/0.0096-1\approx 19.0\%\) greater variance than the 50/50 allocation. The imbalance reduces precision because it leaves one arm with substantially fewer observations.
When Imbalance Can Help
An imbalanced allocation does not always increase the variance. If the arm-specific outcome variances differ, assigning more participants to the arm with greater variance can improve precision compared with equal allocation. This is why the direction of imbalance matters, not just its size.
Worked Example: Checking Both Directions of Imbalance
Consider hypothetical planning probabilities \(p_A=0.50\) and \(p_B=0.10\), with 100 participants total. These give \(v_A=0.50(0.50)=0.25\) and \(v_B=0.10(0.90)=0.09\). Compare 50/50, 62/38, and 38/62 allocations.
For 50/50, the variance is
For 62 participants in arm A and 38 in arm B,
This is lower than 0.006800, so this particular unequal allocation improves precision relative to 50/50 under the stated probabilities. Reversing the allocation gives
This variance is higher than 0.006800. The distinction is that the 62/38 allocation gives more participants to arm A, which has the larger outcome variance. An imbalance in the opposite direction leaves fewer participants in the more variable arm and increases the contrast variance.
The Variance-Minimizing Allocation
The preceding example suggests a general rule under the independent-binomial model, when the total sample size is fixed and each participant has the same cost in either arm. Let \(N=n_A+n_B\). The variance as a function of \(n_A\) is
Differentiate with respect to \(n_A\) and set the derivative to zero:
Rearranging gives \(v_A/n_A^2=v_B/n_B^2\), and hence
This stationary point is a minimum: the second derivative is \(2v_A/n_A^3+2v_B/(N-n_A)^3\), which is positive when both variances and both sample sizes are positive. Thus, when the arm variances are equal, the minimizing allocation is equal. When one arm has greater outcome variance, the variance-minimizing allocation puts more participants in that arm.
For the probabilities in the previous example, the minimizing ratio is \(\sqrt{0.25/0.09}=1.667\), or approximately 62.5 participants in arm A for every 37.5 in arm B when \(N=100\). Integer sample sizes make 62/38 or 63/37 close allocations. The calculation is an idealized planning result: the probabilities, and therefore the variances, may be uncertain.
Worked Example: Estimating the Standard Error From Observed Data
In a hypothetical completed trial, 18 of 40 participants in arm A and 12 of 60 participants in arm B have the binary outcome. Assume independent Bernoulli outcomes within and across arms, with complete outcome data. Calculate the estimated contrast and its estimated standard error.
The sample proportions are \(\hat{p}_A=18/40=0.45\) and \(\hat{p}_B=12/60=0.20\). The estimated difference is \(0.45-0.20=0.25\), or 25 percentage points. Substituting the sample proportions into the standard-error formula gives
Thus, the estimated standard error is about 0.0941, or 9.41 percentage points. The two terms show the arm-specific contributions: about 0.00619 from arm A and 0.00267 from arm B. The calculation estimates uncertainty under the stated model; it does not establish that the model captures every feature of a response-adaptive assignment process.
Interpretation, Diagnostics, and Limitations
The equal-allocation result is a special case, not a rule that every study must split participants equally. If arm variances are equal, equal allocation minimizes the variance of the difference for a fixed total sample size. If they differ, an allocation directed toward the more variable arm may have lower variance. However, choosing that allocation requires credible information about the arm-specific probabilities, and those probabilities may be uncertain before the trial.
A useful diagnostic is to calculate each arm’s contribution, \(v_A/n_A\) and \(v_B/n_B\), or their estimated versions. A much larger contribution from one arm indicates that its observations are limiting precision. Check that the actual arm sizes and outcome counts support the calculation, and examine whether missing outcomes, clustering, or dependence could invalidate the independent-binomial assumptions. Very small counts or a sample proportion of exactly zero or one can also make the plug-in variance estimate unstable; a zero estimated contribution should not be read as proof that the true variance is zero.
The formula has important limits. It concerns a difference in two proportions under a specific model; it does not measure all information in a trial or decide which allocation is ethically or clinically preferable. It also treats arm sizes as fixed and assumes independent observations. In response-adaptive randomization, allocation probabilities may depend on accumulated outcomes, so this simple variance expression is a benchmark rather than a guaranteed analysis formula.
| Allocation choice | Variance implication under the model | Key qualification |
|---|---|---|
| Equal allocation | Minimizes variance when the arm variances are equal. | Equal arm variances are an assumption, not a consequence of equal sample sizes. |
| Unequal allocation toward the more variable arm | Can reduce variance when arm variances differ. | Requires useful estimates or assumptions about those variances. |
| Unequal allocation away from the more variable arm | Can increase variance substantially. | Imbalance alone does not reveal whether precision rises or falls; its direction matters. |
Common Mistake
A frequent error is to claim that every departure from 50/50 allocation increases variance. That conclusion is correct only under conditions such as equal arm variances and a fixed total sample size. With unequal variances, an allocation that gives more participants to the more variable arm can have lower variance than 50/50, as the 62/38 calculation above shows. Always substitute the arm sizes and arm-specific variances into the formula before making a comparison.
Check Your Understanding
Try each question, then check the solutions below.
- With \(N=80\) and equal arm variances \(v_A=v_B=0.20\), calculate the variance under 40/40 allocation and under 50/30 allocation.
- If \(v_A=0.16\) and \(v_B=0.36\), which arm should receive more participants to minimize variance under equal per-participant costs?
- Why is the variance of a difference between independent sample proportions the sum of the two arm variances?
- What does a plug-in estimated variance of zero in one arm mean, and why should it be interpreted cautiously?
Solutions. (1) For 40/40, \(0.20/40+0.20/40=0.0100\). For 50/30, \(0.20/50+0.20/30=0.0040+0.006667=0.01067\), so the unequal allocation has greater variance. (2) Arm B has the larger variance, so it should receive more participants; the minimizing ratio is \(n_A/n_B=\sqrt{0.16/0.36}=2/3\). (3) Independence makes the covariance zero, so \(\operatorname{Var}(\hat{p}_A-\hat{p}_B)=\operatorname{Var}(\hat{p}_A)+\operatorname{Var}(\hat{p}_B)\). (4) It means the observed proportion is zero or one, making the plug-in estimate \( \hat{p}(1-\hat{p})/n\) equal to zero. With limited data, that does not prove the underlying outcome probability is zero or one, so the estimated uncertainty may be unstable.