Simulation With Unequal Allocation Ratios

Why Allocation Ratios Matter in MAMS Simulation

A multi-arm multi-stage (MAMS) design can allocate participants equally across arms, or assign different numbers to control and experimental arms. More participants on control may improve the precision of several treatment-control comparisons at once, but it leaves fewer participants in each experimental arm when the total sample size is fixed. Unequal treatment-arm sizes can also alter comparisons between experimental treatments. These choices affect the information available at each interim analysis and, consequently, simulated power, error rates, and the number of participants expected to enter the trial.

Simulation is useful because these operating characteristics depend on the whole design: allocation by stage, shared-control comparisons, interim decisions, and the stopping rules. A key principle is that changing allocation does not automatically invalidate a correctly calibrated test. It does change the sampling variances and dependence structure, so the simulation must represent the allocation actually intended and use decision rules appropriate to it.

Define the Allocation Before Simulating

A ratio alone is not a complete design description. For example, “2:1” does not say whether every treatment arm has twice as many participants as control, whether the ratio changes across stages, or whether total enrollment is fixed. A simulation specification should state the planned arm counts, how randomization achieves them, and whether totals refer to participants enrolled or participants with observed outcomes. In the examples below, outcomes are complete and the stated sample sizes are the numbers analyzed.

To isolate allocation effects, suppose first that outcomes are independent between participants, have a common variance \(\sigma^2\), and are normally distributed within each arm. For a treatment-control mean difference, \(\bar{x}_i-\bar{x}_0\), the variance is

This formula assumes independent participants and equal outcome variance in the two arms. With a suitable large-sample approximation, it also describes the variance of many estimated treatment effects beyond normally distributed outcomes. If variances differ, the corresponding arm-specific variances must be used rather than assuming a common \(\sigma^2\). In a MAMS simulation, each look uses the sample sizes accrued by that look; using final sample sizes to represent an earlier interim analysis gives the wrong information.

What a Fixed Total Does to Precision

For one comparison with fixed total \(N=n_0+n_i\) and ratio \(r=n_i/n_0\), the arm sizes are \(n_0=N/(1+r)\) and \(n_i=rN/(1+r)\). Substitution into the variance formula gives

With equal allocation, \(r=1\), this variance is \(4\sigma^2/N\). At the same total \(N\), the variance relative to equal allocation is \((1+r)^2/(4r)\). This quantity is at least 1, since \((1+r)^2-4r=(r-1)^2\geq 0\). Thus, for a single comparison with equal variances and equal per-participant costs, equal allocation minimizes the variance when total sample size is fixed. That result does not imply equal allocation is always best in a multi-arm trial: a control participant can contribute to several comparisons, and the design may value those comparisons differently.

Shared Control and Unequal Treatment Arms

When several experimental treatments share a control group, the treatment-control estimates are correlated because they contain the same estimated control mean. Under the independent, equal-variance model, for two experimental arms \(i\) and \(j\),

The covariance is positive: a control mean that happens to be high shifts both treatment-control differences downward. Changing the control allocation changes both the variance of each contrast and this covariance. Therefore, a simulation that applies separate one-comparison calculations but ignores the shared-control structure can misrepresent joint outcomes, such as the chance that at least one treatment succeeds or that all specified treatments succeed.

Unequal allocation among experimental arms has its own consequences. A smaller treatment arm generally has a less precise treatment-control contrast and may have lower power, especially if the treatment effect of interest is similar across arms. A treatment-treatment contrast, \(\bar{x}_i-\bar{x}_j\), has variance \(\sigma^2(1/n_i+1/n_j)\); it depends directly on both experimental-arm sizes, not on the control size. Hence, adding participants to control can be a reasonable trade-off for treatment-control comparisons while making treatment-treatment comparisons less precise if it reduces both active-arm sizes.

Allocation and Simulated Operating Characteristics

For each proposed allocation, simulate complete trial replicates under the scenarios required by the design question. Under a global null scenario, all treatment effects are null; the proportion of replicates with one or more false treatment declarations estimates the family-wise error rate. Other null configurations can assess error rates when some treatments are effective and others are not. Under alternative scenarios, the proportion of replicates meeting a specified success criterion estimates power. Define the criterion explicitly: a pairwise power, a probability of at least one success, and a probability of success for every target treatment are different quantities.

For a simulated event proportion \(\hat{p}\) from \(R\) independent replicates, the Monte Carlo standard error is \(\sqrt{\hat{p}(1-\hat{p})/R}\). This measures simulation noise, not uncertainty about the clinical effect or the validity of the model. The earlier material on Monte Carlo standard errors and confidence intervals for simulated error rates gives the general interpretation; here the essential comparison is that estimates from different allocations have their own simulation uncertainty.

Unequal allocation can also change expected sample size in a multi-stage trial. The maximum planned sample size may be identical across allocation schemes, yet the expected total differs because the probability of stopping early or continuing to later stages can change. Under a particular alternative, lower power at an interim analysis may lead more replicates to continue; under another scenario, changes in the joint evidence across arms can alter early stopping in the opposite direction. The direction is design- and scenario-specific, so it should be estimated rather than assumed.

A Useful Allocation Calculation—and Its Limits

A simple calculation can help choose a starting allocation for several equally important treatment-control comparisons. Suppose there are \(k\) experimental arms, each with \(n_t\) participants, \(n_0\) control participants, fixed total \(N=n_0+kn_t\), equal variances, equal per-participant costs, and the objective is to minimize the sum of the \(k\) treatment-control variances. Ignoring the common factor \(\sigma^2\), that sum is \(k(1/n_t+1/n_0)\). Substituting \(n_t=(N-n_0)/k\), the quantity to minimize is

Differentiating with respect to \(n_0\) and setting the derivative to zero gives \(k/(N-n_0)^2=1/n_0^2\). The positive solution is \(N-n_0=\sqrt{k}\,n_0\), so \(n_0=N/(1+\sqrt{k})\) and \(n_0/n_t=\sqrt{k}\). This is a result for that specific objective and set of assumptions, not a universal MAMS allocation rule. It does not directly optimize treatment-treatment comparisons, power across unequal effects, error rates, recruitment constraints, or costs that differ by arm.

Simulation Workflow, Diagnostics, and Pitfalls

A defensible allocation comparison keeps the design and simulation aligned. Use the following sequence for each proposed allocation:

Diagnostics should include checking that simulated arm counts match the allocation plan, that interim sample sizes are cumulative as intended, and that the variance of simulated treatment-control estimates is consistent with \(\sigma^2(1/n_i+1/n_0)\) under the simple model. Inspect error rates under the relevant null configurations, not only under a single global null. If simulated error rates depart materially from the target, first check the analysis and boundary calibration for that allocation before attributing the difference to the ratio itself.

A common mistake is to assume that doubling the treatment-to-control ratio must double the information for that treatment. At a fixed total, the ratio changes both arm sizes, and the contrast variance depends on the reciprocal of each size. The correction is to calculate or simulate the contrast precision using the actual \(n_i\) and \(n_0\), then assess all arms and stages together. Another mistake is to compare simulated event proportions without accounting for Monte Carlo variation. A small numerical difference between two estimates may simply reflect finite replicate counts.

The main alternatives are equal allocation, a common control-heavy allocation for all treatments, and unequal ratios tailored to arm-specific priorities. Equal allocation is efficient for a single equal-cost two-arm comparison at fixed total. A larger shared control can be useful when several treatment-control comparisons are central, but it can reduce the precision of treatment-treatment comparisons or of particular treatment arms. No ratio should be chosen on the basis of one operating characteristic alone: consider error control, the required power for each relevant comparison, expected and maximum sample size, and feasibility.