Operating Characteristics Under Unequal Allocation

Why Allocation Must Be Evaluated at a Fixed Total

A treatment-to-control allocation ratio determines how a fixed number of participants is divided between trial arms. A 1:1 design assigns roughly half to each arm; a 2:1 design assigns roughly two participants to treatment for every one assigned to control. Unequal allocation may be attractive when more treatment experience is valuable, but under a fixed total enrollment it leaves fewer participants in the other arm. The statistical consequence is not captured by the ratio alone: it depends on the outcome, treatment effect, decision rule, and data-generating assumptions.

Operating characteristics are repeated-trial summaries. For example, under a specified true treatment effect and outcome model, the probability that a trial meets a prespecified Bayesian decision rule is a decision probability or power for that scenario. It is calculated by repeatedly generating trial data, fitting the planned analysis, and recording whether the rule is met. It is not the posterior probability from one trial. A fair allocation comparison changes the allocation while keeping total enrollment, outcome-generating assumptions, analysis, prior, and decision rule fixed.

Allocation and Precision at Fixed Enrollment

Let \(N\) be total enrollment, \(n_T\) the number assigned to treatment, and \(n_C\) the number assigned to control, so \(N=n_T+n_C\). Define the allocation ratio \(r=n_T/n_C\). Ignoring rounding for the moment, the arm sizes are:

For a continuous outcome with independent observations and a common known variance \(\sigma^2\) in both arms, the sample mean difference \(D=\bar{X}_T-\bar{X}_C\) estimates the treatment-minus-control mean difference. Its sampling variance and standard error are:

This expression assumes independent outcomes, the stated common variance, and no additional source of dependence such as clustering. The calculation is an illustration of the allocation trade-off, not a substitute for simulating the planned trial when its outcome model or analysis is more complex.

At fixed \(N\), the balanced allocation minimizes this variance under the equal-variance assumptions. The key identity is:

Since \((n_T-n_C)^2\geq 0\), we have \(4n_Tn_C\leq(n_T+n_C)^2=N^2\), with equality only when \(n_T=n_C\). For integer arm sizes, \(n_Tn_C=\{N^2-(n_T-n_C)^2\}/4\), so the product is maximized when \(|n_T-n_C|\) is as small as possible: 0 for even \(N\) and 1 for odd \(N\). Thus the closest-to-equal allocation makes \(n_Tn_C\) largest, and \(N/(n_Tn_C)\), and therefore the variance, smallest. The result does not say that balanced allocation is best for every clinical objective. It says that, for this particular two-arm mean-difference precision criterion and these assumptions, unequal allocation costs precision at fixed total enrollment.

Compare Operating Characteristics, Not Just Standard Errors

The standard-error calculation is a useful diagnostic for a continuous, equal-variance setting. A trial-calibration simulation should nevertheless evaluate the decision rule actually planned for the trial. For example, let \(\theta\) denote the treatment-minus-control effect, and suppose the rule declares efficacy when the posterior probability that \(\theta>0\) is at least \(\gamma\). Under a specified true value of \(\theta\), the operating characteristic is the repeated-trial probability of crossing that threshold. Changing allocation can alter the distribution of the data and therefore the probability of crossing the rule.

For a clean comparison, both designs should use the same total enrollment and the same assumptions about outcomes, true effect, prior, analysis, decision threshold, and handling of trial events. If allocation is randomized, its implementation should match the simulation. If arm counts must be integers, specify the rounding rule and use the resulting counts in both the data generation and analysis. Record the decision indicator for every replicate; the estimated operating characteristic is the proportion of replicates meeting the rule.

Designing and Checking the Simulation

The simulation should make allocation the only planned design change. A practical sequence is:

For an estimated probability \(\hat{q}\) based on \(M\) independent replicates, the Monte Carlo standard error is approximately \(\sqrt{\hat{q}(1-\hat{q})/M}\). This describes simulation error in estimating the operating characteristic; it is not the standard error of an individual trial's treatment-effect estimate. If both designs are evaluated using deliberately paired random-number streams, their estimates may be correlated. In that case, uncertainty for the difference should account for the pairing rather than automatically using the independent-runs formula.

Diagnostics include verifying that arm sizes sum to \(N\), checking that simulated outcomes have the intended distributions within each arm, and confirming that the same analysis code and threshold are used for both designs. Compare simulated treatment-effect means and standard deviations with their expected values; for the continuous example, the treatment-effect estimate should average near the true effect and have a standard deviation near the calculated \(SE(D)\). Also check that simulation estimates stabilize as the number of replicates increases. A difference that shrinks or changes direction substantially with more replicates may have been Monte Carlo noise rather than a design effect.

Comparisons, Limitations, and Common Pitfalls

Balanced allocation is a useful reference because it minimizes the variance of a two-arm mean difference at fixed total enrollment when outcome variances are equal. Unequal allocation may still be preferred for clinical, ethical, recruitment, or treatment-exposure reasons. Its operating characteristics should be assessed against those objectives, not rejected solely on the basis of a standard-error formula. If arm variances differ, the variance of the mean difference is \(\sigma_T^2/n_T+\sigma_C^2/n_C\); the equal-variance result no longer directly describes the comparison. Other outcome types and analyses likewise require their own simulation model.

A common mistake is to compare, for example, 60:60 with 80:40 and conclude that allocation caused a difference when the unequal design also used a larger total sample, a different prior, or a different decision threshold. That comparison changes several factors at once. The correction is to hold total enrollment and all outcome-generating and analysis assumptions fixed, change only the allocation, and report Monte Carlo uncertainty for the estimated difference. A second mistake is to interpret a lower simulated efficacy probability as evidence that the posterior probability in an individual trial is lower by the same amount. The former is a repeated-trial frequency; the latter is a data-dependent posterior quantity. Keep those interpretations distinct.

Finally, a single scenario does not characterize all trial performance. An allocation that performs less well under one effect and outcome variance may behave differently in another scenario. The conclusions here are conditional on the scenarios simulated. Do not extrapolate them to unexamined assumptions, or treat a small numerical difference as meaningful if it is comparable to Monte Carlo uncertainty.