Unequal Allocation in Two-Sample Designs

Why Allocation Matters

A two-sample study need not place the same number of observations in each group. If one group is more variable than the other, assigning it more observations can improve the precision of the estimated difference in means. The benefit depends on both the total sample size and the allocation ratio; an unequal split is not automatically more efficient.

The calculations below concern a difference in two independent population means. They show how allocation affects standard error and power, and how to select an allocation when the variances are known or have reasonable planning estimates. These results are aligned with the SOA VEE Mathematical Statistics syllabus.

The group labels matter when interpreting \(r\): \(r=2\) means twice as many observations in group 2 as in group 1. For a fixed total \(N\), the corresponding sample sizes are \(n_1=N/(1+r)\) and \(n_2=Nr/(1+r)\). The standard error of the difference is the square root of its variance.

Allocation, Standard Error, and Power

For a two-sided test of \(H_0:\mu_1-\mu_2=0\) at significance level \(\alpha\), suppose observations are independent and normally distributed with known variances. The statistic \(Z=(\bar{X}_1-\bar{X}_2)/\sqrt{\sigma_1^2/n_1+\sigma_2^2/n_2}\) has a standard normal distribution under the null hypothesis. If the true difference is \(\Delta=\mu_1-\mu_2\), its distribution has mean \(\lambda=\Delta/\sqrt{\sigma_1^2/n_1+\sigma_2^2/n_2}\) and variance 1. Thus the allocation affects power by changing the standard error and therefore the noncentral mean \(\lambda\).

Here \(\Phi\) is the standard normal cumulative distribution function and \(z_q\) is its \(q\)-quantile. The expression adds the probabilities of falling below the lower critical value and above the upper critical value. For a positive \(\Delta\), the upper-tail probability is usually the larger one, but the lower-tail probability is part of the two-sided test and should not be omitted.

When the variances are unknown, a pooled two-sample \(t\) test has an exact noncentral-\(t\) power calculation under independent normal populations with a common variance. Its degrees of freedom are \(N-2\), and its noncentrality parameter is \(\Delta/[\sigma\sqrt{1/n_1+1/n_2}]\). That common-variance assumption is important: if the population variances differ, the pooled test is not generally justified. A Welch test or a normal planning approximation based on separate variance estimates may be more appropriate. The normal formula above is exact for known variances and normal observations; with estimated variances or nonnormal observations it is an approximation, whose adequacy depends on the model and sample sizes.

The Variance-Minimizing Allocation

For a fixed total sample size and equal cost per observation, the allocation that minimizes the variance of the difference in means gives more observations to the group with the larger standard deviation. To see why, write the variance in terms of the ratio \(r=n_2/n_1\), using \(n_1=N/(1+r)\) and \(n_2=Nr/(1+r)\).

The optimal ratio and the resulting minimum variance are \(n_2/n_1=\sigma_2/\sigma_1\) and \(\min V=(\sigma_1+\sigma_2)^2/N\). Equivalently, \(n_1/n_2=\sigma_1/\sigma_2\). The standard deviations, not the variances, determine the ratio. If \(\sigma_1=\sigma_2\), this result reduces to equal allocation.

Planning Total Sample Size

A common planning question specifies a target power \(1-\beta\), a two-sided significance level \(\alpha\), a meaningful difference \(|\Delta|\), and planning values for the standard deviations. Under the same independent, known-variance normal model, the optimal continuous allocation makes the standard error \((\sigma_1+\sigma_2)/\sqrt{N}\). A standard normal approximation chooses the noncentral mean to be about \(z_{1-\alpha/2}+z_{1-\beta}\). Solving for \(N\) gives the following planning formula.

To avoid ambiguity in using the formula, the optimal allocation is \(n_1:N\) and \(n_2:N\) in the proportions \(n_1/N=\sigma_1/(\sigma_1+\sigma_2)\) and \(n_2/N=\sigma_2/(\sigma_1+\sigma_2)\). The total from the planning formula is continuous, so round the total sample size up and choose integer group sizes near those proportions. Then recompute the standard error and power using the actual integers. The formula is an approximation; the final check is particularly important when the required sample size is small or the model calls for a \(t\) test instead.

Interpretation and Exam Traps

The allocation result is about precision under a specified model, not a universal instruction to make group sizes unequal. With equal variances and equal costs, balance is optimal. When variances differ, the higher-variance group generally needs more observations to contribute a similar amount of uncertainty to the estimated difference.

A frequent error is to reverse the optimal ratio. For example, if \(\sigma_1=4\) and \(\sigma_2=8\), the optimal ratio is \(n_2/n_1=8/4=2\), not \(1/2\). Assigning twice as many observations to group 1 would instead give \(n_1=60,n_2=30\) at \(N=90\); its variance would be \(16/60+64/30=0.2667+2.1333=2.4000\), larger than the optimal allocation's variance of \(1.6000\). The correction is to match the group-size ratio to the ratio of standard deviations in the same group order.

Other assumptions and practical limits matter. The formulas presume independent observations and a meaningful total sample size. If observations are paired, the variance of a paired difference is the relevant quantity instead. If group costs differ, the equal-cost optimum may waste resources. If the standard deviations are estimated from prior information, they should be treated as planning values and sensitivity to plausible alternatives should be considered. Finally, when using a pooled \(t\) test, unequal population variances violate its common-variance assumption; changing the allocation does not repair that assumption.