Why Unequal Variances Change Information
For a two-group mean comparison, precision depends on more than the total number of participants. If outcomes vary more in one group than the other, adding a participant to that group generally contributes less precision than adding one to the less variable group. This matters when translating enrollment into information for a group sequential design: the same total enrollment can correspond to different information depending on the group allocation and the two outcome variances.
Let group A have \(n_A\) independent observations with population mean \(\mu_A\) and variance \(\sigma_A^2\), and group B have \(n_B\) independent observations with population mean \(\mu_B\) and variance \(\sigma_B^2\). The parameter of interest is the mean contrast \(\theta=\mu_A-\mu_B\), estimated by \(\bar{x}_A-\bar{x}_B\). Assume observations are independent within and between groups, with a common mean and variance within each group. The variances may differ across groups. For the exact normal distribution of the contrast, also assume normal outcomes in each group; with adequate sample sizes, a normal approximation may instead be appropriate under standard central limit conditions.
The variance formula follows because the sample means are independent and have variances \(\sigma_A^2/n_A\) and \(\sigma_B^2/n_B\). Subtracting independent estimates adds their variances. If \(H_0:\mu_A-\mu_B=0\) is tested with known variances, the standardized contrast is \(Z=(\bar{x}_A-\bar{x}_B)/\sqrt{V}\). Under the normal model and \(H_0\), \(Z\sim N(0,1)\). The same \(V\) determines the standard error and information; the observed difference itself does not change the information.
Deriving the Contrast Information
Write the two sample means as averages of their observations. Independence gives \(\operatorname{Var}(\bar{x}_A)=\sigma_A^2/n_A\) and \(\operatorname{Var}(\bar{x}_B)=\sigma_B^2/n_B\). Because the covariance between the two means is zero,
This is the justification for the information definition: for this contrast, the inverse variance is the natural measure of precision. If variances are estimated rather than known, the estimated standard error uses the separate within-group variances, not a common pooled variance. The earlier material on information when variance is estimated treats the additional issue that estimated variances make the information itself uncertain.
Worked Example: Computing Information With Unequal Variances
Suppose a hypothetical study compares a continuous outcome between two independent groups. Assume normal outcomes with known standard deviations of 10 points in group A and 15 points in group B, with \(n_A=25\) and \(n_B=36\). Calculate the variance, standard error, and information for the difference in means.
The group variances are \(\sigma_A^2=10^2=100\) points squared and \(\sigma_B^2=15^2=225\) points squared. Therefore,
The standard error is \(\sqrt{10.25}=3.202\) points. The information is \(I=1/10.25=0.09756\) per point squared. The calculation checks the assumptions used: the groups are specified as independent, the variances are treated as known, and normal outcomes are assumed for exact normal inference. The larger variance in group B contributes \(6.25\) points squared to the contrast variance, compared with \(4\) from group A, despite group B having more observations.
How Allocation Interacts With the Variances
When total enrollment \(N=n_A+n_B\) is fixed, the allocation affects the two contributions to \(V\). For continuous sample sizes, the variance is minimized by allocating in proportion to the group standard deviations, not the group variances. To see this, set \(n_B=N-n_A\) and minimize \(V(n_A)=\sigma_A^2/n_A+\sigma_B^2/(N-n_A)\) over \(0<n_A<N\). The derivative is \(-\sigma_A^2/n_A^2+\sigma_B^2/(N-n_A)^2\). Setting it to zero gives \(\sigma_A/n_A=\sigma_B/n_B\), hence \(n_A/n_B=\sigma_A/\sigma_B\). The second derivative is positive throughout the feasible interval, so this stationary point is the unique continuous minimum.
This result assumes there are no group-specific costs, caps, or other allocation constraints and that the variances are adequately known for planning. It is an optimum for continuous allocation. Actual participants are counted in integers, so the formula supplies a target ratio, not necessarily an attainable allocation. When counts must be integers, evaluate feasible allocations near the continuous optimum.
Worked Example: Choosing an Integer Allocation
A hypothetical trial has a fixed total of \(N=40\) participants. Planning standard deviations are 4 outcome units in group A and 8 in group B. Find the continuous target and the best nearby integer allocation, then compare with equal allocation.
The continuous target ratio is \(n_A/n_B=4/8=1/2\), so \(n_A=40(4/12)=13.333\) and \(n_B=26.667\). These fractional counts cannot be used in the trial. The two adjacent feasible allocations are (13, 27) and (14, 26). Since \(V(n_A)\) is strictly convex and decreases up to the continuous minimum then increases, one of these adjacent counts is the integer minimum.
For (13, 27), \(V=4^2/13+8^2/27=16/13+64/27=3.6011\) outcome units squared. For (14, 26), \(V=16/14+64/26=3.6044\) outcome units squared. Thus (13, 27) has the lower variance of the two. Equal allocation gives \(V=16/20+64/20=4.000\) outcome units squared. The integer target therefore slightly improves on equal allocation, while the continuous minimum \(V_{\min}=12^2/40=3.600\) is a lower bound that the integer design cannot attain exactly.
The allocation result is also a useful way to interpret marginal value. At the continuous optimum, the two groups satisfy \(\sigma_A/n_A=\sigma_B/n_B\); the reduction in variance from shifting a small amount of sample size between groups is balanced. If allocation is fixed by design or operational constraints, there is no need to force this optimum. The correct information is still calculated from the actual \(n_A,n_B,\sigma_A^2,\sigma_B^2\).
Information Fractions During Accrual
At an interim analysis, let the accrued sample sizes be \(n_{A,t}\) and \(n_{B,t}\). Under stable group variances and independent outcomes, interim information is \(I_t=1/(\sigma_A^2/n_{A,t}+\sigma_B^2/n_{B,t})\). If the planned final sample sizes are \(n_{A,T}\) and \(n_{B,T}\), the corresponding information fraction is \(I_t/I_T\). This ratio need not equal the fraction of total enrollment accrued: the two groups can contribute differently to the contrast variance, and their interim accrual proportions may differ from their planned final proportions.
Worked Example: An Interim Fraction With Unequal Accrual
In a hypothetical two-group trial, assume stable standard deviations of 5 outcome units in group A and 8 in group B. At an interim analysis, 10 participants have accrued in A and 30 in B. The planned final counts are 40 in each group. Calculate the information fraction and compare it with the enrollment fraction.
At interim, the contrast variance is \(V_t=25/10+64/30=2.5+2.1333=4.6333\) outcome units squared, so \(I_t=1/V_t\). At the planned final analysis, \(V_T=25/40+64/40=0.625+1.600=2.225\) outcome units squared, so \(I_T=1/V_T\). Taking the ratio without rounding the intermediate information values gives
Thus interim information is about 48.0% of planned final information, whereas 40 of the planned 80 participants have accrued, an enrollment fraction of 50%. The difference arises from the interim allocation: only one quarter of the final group A count has accrued, compared with three quarters of the final group B count. The higher-variance group B contributes more variance per participant, so its greater interim accrual does not translate directly into the enrollment fraction as information.
Assumptions, Diagnostics, and Alternatives
The calculation is only as credible as its variance and sampling assumptions. Independence matters: within-person pairing, repeated measurements, clustering, or other dependence changes the variance formula. The earlier material on paired comparisons, repeated measurements, unequal cluster sizes, and overdispersion addresses settings in which a simple independent-observation calculation is not appropriate. Here the target is specifically a contrast of two independent group means with potentially different within-group variances.
In practice, inspect group-specific outcome variability and the accrued counts separately. A single overall standard deviation can conceal a substantially more variable group. Also ask whether the planned standard deviations remain plausible as participants accrue, and whether missing outcomes or changes in measurement could alter the effective sample sizes or variance. If group variances are estimated from small samples, the plug-in information can be unstable; it should not be treated as known merely because a formula produces a precise-looking number.
| Approach | Variance basis | When it is appropriate |
|---|---|---|
| Unequal-variance contrast | \(\sigma_A^2/n_A+\sigma_B^2/n_B\) | Independent groups whose variances may differ |
| Pooled-variance contrast | A common variance estimate multiplied by \(1/n_A+1/n_B\) | When the equal-variance model is justified |
| Paired or dependence-adjusted contrast | Includes covariance or dependence structure | When observations are paired, repeated, or clustered |
A common mistake is to use a pooled variance simply because the two groups are being compared, or to treat total enrollment as information. The correction is to specify the sampling model first and calculate each group's contribution, \(\sigma_g^2/n_g\), before taking the reciprocal. Another mistake is to report the continuous optimum as an exact participant allocation. It is a planning target; integer counts must be checked among feasible allocations. Finally, a calculated information fraction is not itself evidence that variance assumptions hold: a substantial change in group variability can make the planned fraction a poor representation of realized precision.
Check Your Understanding
Answer each question, then compare with the short solutions below.
- With \(\sigma_A=3\), \(\sigma_B=6\), \(n_A=18\), and \(n_B=24\), calculate the contrast variance and information.
- For a fixed total and unconstrained continuous allocation, what sample-size ratio minimizes the variance if \(\sigma_A=2\sigma_B\)?
- Why can an interim information fraction differ from the interim enrollment fraction?
- Name one assumption that would make the independent two-group variance formula inappropriate.
Solutions. (1) \(V=9/18+36/24=0.5+1.5=2.000\) outcome units squared, so \(I=1/2=0.5000\) per squared outcome unit. (2) The target ratio is \(n_A/n_B=\sigma_A/\sigma_B=2\). (3) The group-specific variance contributions and interim allocation determine the contrast variance, so information need not grow in proportion to total enrollment. (4) Dependence, such as pairing, repeated observations, or clustering, invalidates the assumption that the two sample means have the stated independent variances.