Effect Of Allocation Ratio On Required Sample Size

Why Allocation Ratio Changes the Required Sample Size

A fixed allocation ratio determines how a trial’s total sample is divided between treatment and control. When outcome variances and the treatment effect are held constant, changing that division changes the precision of the estimated difference in means. The practical question is how many participants are needed at each ratio to attain a specified power.

The key comparison is between designs that use the same outcome model, treatment effect, significance level, and target power. If any of those inputs changes between calculations, the resulting sample sizes no longer isolate the effect of allocation ratio. The calculations below use a normal approximation for a two-sided test; at small sample sizes, a calculation based on the appropriate finite-sample test may be preferable.

For example, \(r=2\) means two treatment participants for every control participant. It does not mean that treatment receives two-thirds more participants than control: it receives twice as many. In every comparison, the ratio’s direction must be stated, because reversing the numerator and denominator changes its interpretation.

Model, Assumptions, and Sample-Size Formula

Suppose independent participants in the two groups have continuous outcomes with means \(\mu_T\) and \(\mu_C\), and variances \(\sigma_T^2\) and \(\sigma_C^2\). The treatment effect to detect is \(\delta=\mu_T-\mu_C\), with \(|\delta|>0\). For planning, assume these variances and the effect are fixed and known inputs. The usual large-sample normal approximation for the difference in sample means has standard error \(SE=\sqrt{\sigma_T^2/n_T+\sigma_C^2/n_C}\).

The test is two-sided, with null hypothesis \(H_0:\mu_T-\mu_C=0\), alternative hypothesis \(H_a:\mu_T-\mu_C\ne0\), and significance level \(\alpha\). The normal approximation treats the standardized test statistic as having variance 1 and, under the planning alternative, mean \(|\delta|/SE\). This setup requires independent observations within and between groups, a meaningful variance model, and an analysis consistent with the planned comparison. In small samples, the normal approximation may not describe the test’s power well.

Write \(z_q\) for the \(q\)-quantile of the standard normal distribution. The usual planning approximation achieves target power \(1-\beta\) when the standardized effect is at least \(z_{1-\alpha/2}+z_{1-\beta}\). Substituting the allocation ratio into the standard error gives a formula for the total sample size.

When both groups have the same variance, \(\sigma_T^2=\sigma_C^2=\sigma^2\), this simplifies to \(N\approx (z_{1-\alpha/2}+z_{1-\beta})^2\sigma^2(1+r)^2/(r\delta^2)\). The allocation-dependent factor is then \((1+r)^2/r=r+2+1/r\). It is smallest at \(r=1\), so equal allocation requires the smallest total sample under this equal-variance model when the cost and other conditions are the same.

Why the Formula Has This Shape

The approximation can be justified directly from the test’s power. Under a positive effect, the standardized statistic is approximately normal with mean \(\lambda=\delta/SE\) and variance 1. The two-sided test rejects above \(z_{1-\alpha/2}\) or below \(-z_{1-\alpha/2}\). Its approximate power is therefore \(P(Z>z_{1-\alpha/2})+P(Z<-z_{1-\alpha/2})\) when \(Z\) has that shifted normal distribution.

For a positive effect, the upper-tail rejection probability is \(\Phi(\lambda-z_{1-\alpha/2})\). Setting this probability approximately equal to \(1-\beta\) gives \(\lambda\approx z_{1-\alpha/2}+z_{1-\beta}\). This is the familiar sample-size approximation; it neglects the small lower-tail rejection probability under a positive alternative. Replacing \(SE^2\) by the allocation-ratio expression and solving for \(N\) yields the displayed formula. The omitted tail is one reason to treat the result as a planning approximation, not an exact power calculation.

With equal variances, the allocation factor is \(r+2+1/r\). Its derivative with respect to \(r\) is \(1-1/r^2\), which is negative for \(0<r<1\), zero at \(r=1\), and positive for \(r>1\). Thus the factor decreases up to equal allocation and increases beyond it. This shows why imbalance raises the required total sample in this setting.

Comparing Several Ratios

A useful way to compare candidate ratios is to keep every input except \(r\) unchanged. The next example does this for a hypothetical trial with a common outcome standard deviation in both groups.

Checking the Calculation and Its Limits

Before using a sample-size comparison, check that the planned effect, variances, significance level, and power target are identical across ratios. Confirm that \(r\) is treatment divided by control, and verify that the resulting group sizes satisfy \(n_T/n_C\approx r\). A frequent arithmetic error is to calculate \(n_T\) and \(n_C\) from a rounded total and then report a ratio inconsistent with those group sizes.

Another diagnostic is the equal-variance allocation factor. It must be at least 4, since \(r+2+1/r\ge4\) for positive \(r\). If a calculation at \(r\ne1\) gives a smaller factor than 4, check the ratio definition and arithmetic. Also check units: \(\sigma^2\) and \(\delta^2\) both have outcome-units squared, so their ratio and the resulting sample size are dimensionless counts.

The formula assumes complete, independent observations and the specified variance and effect. Missing outcomes, clustering, unequal costs per participant, or a different analysis can change the required enrollment. If outcomes are not adequately represented by the assumed large-sample model, the approximation may give inaccurate power; use a calculation matched to the planned analysis, especially for small samples. Integer rounding can also shift the exact power slightly.

A common mistake is to argue that a 2:1 allocation should need fewer total participants because more people receive the treatment. That reasoning ignores the loss of precision caused by having fewer control observations. With equal variances, precision depends on \(1/n_T+1/n_C\); for a fixed total, that sum is minimized by equal group sizes. Unequal allocation may still be chosen for reasons outside this calculation, but it does not reduce total sample size for the same target power under the stated equal-variance assumptions.

These calculations compare fixed allocation designs. They do not establish the sample size for a response-adaptive design, whose realized group sizes may vary with observed outcomes. Nor does the formula by itself determine which ratio is best when practical or ethical considerations also matter. It isolates one question: how the fixed division of a given total affects the approximate sample size required under the stated model.