Allocating Subjects Across Crossover Sequences

Why Sequence Allocation Matters

In a two-period, two-treatment crossover study, each participant receives both treatments, but in a different order. The two sequences are test then reference (TR) and reference then test (RT). Allocation determines how many participants contribute information under each order. With a fixed total sample size, an imbalance generally reduces precision because the treatment contrast depends on information from both sequences.

The goal is not simply to make the sequence counts look alike. The allocation should support an estimable treatment contrast, retain enough participants in each sequence, and reflect plausible differences in completion or measurement variability. The discussion below uses a complete-data, two-sequence design and log-transformed pharmacokinetic outcomes, such as log AUC or log Cmax. A treatment difference on this scale corresponds to a log test-to-reference ratio.

The standard calculation assumes randomized sequence assignment; a common treatment effect and period effect; no differential carryover affecting the treatment contrast; complete outcomes in both periods; and independent participants. For the basic model, assume the within-participant residuals in the two periods are independent and normally distributed with common variance \(\sigma_w^2\). These assumptions give a clear benchmark for comparing allocations. Departures such as missing periods, sequence-dependent variability, or carryover require additional analysis rather than a mechanical use of the benchmark.

Deriving the Precision Cost of Imbalance

Write \(D_i=Y_{i1}-Y_{i2}\) for participant \(i\)'s first-period outcome minus second-period outcome. Let \(\tau\) denote the test-minus-reference treatment effect and \(\lambda\) the period-1-minus-period-2 effect. Under the stated model, the expected within-participant difference is \(\tau+\lambda\) in the TR sequence and \(-\tau+\lambda\) in the RT sequence. The shared period effect cancels when the two sequence means are contrasted.

For each sequence, the variance of a within-participant difference is \(2\sigma_w^2\), since it is the difference of two independent residuals, each with variance \(\sigma_w^2\). The sequence groups contain different participants, so their sample means are independent. Therefore:

Replacing \(\sigma_w^2\) with its estimate gives the standard error. If \(s_D^2\) is the pooled sample variance of the \(D_i\) values across the two sequence groups, then \(s_D^2\) estimates \(2\sigma_w^2\), and:

A short optimization shows why equal allocation is preferred under equal variances. For a fixed \(N\), write \(n_{TR}=pN\) and \(n_{RT}=(1-p)N\), where \(0<p<1\). The allocation-dependent part of the variance is proportional to \(1/[p(1-p)]\). Since \(p(1-p)\leq 1/4\), with equality only at \(p=1/2\), variance is minimized when the counts are equal. Relative to equal allocation, the variance multiplier is \(1/[4p(1-p)]\), and the standard-error multiplier is its square root.

Degrees of Freedom and the Equivalence Test

Under the classical complete-data analysis, estimate the common within-participant variance by pooling the within-sequence variability in \(D_i\). Each sequence mean uses one degree of freedom, so the pooled variance has \(N-2\) degrees of freedom. The treatment statistic is \(\widehat{\tau}/SE(\widehat{\tau})\), compared with a \(t\) distribution with \(N-2\) degrees of freedom under the null treatment contrast.

This point is easy to miss: with all participants evaluable and both sequences represented, changing 20:20 to 24:16 does not change the classical residual degrees of freedom when \(N\) remains 40. It changes the standard error, not the count \(N-2=38\). If participants are lost, the evaluable total and hence the residual degrees of freedom may fall. If only one sequence remains, the treatment and period effects cannot be separated by this contrast. Analyses using different variance assumptions or small-sample degrees-of-freedom approximations can produce different degrees of freedom; those results should not be substituted for the classical model without stating the analysis.

For average bioequivalence, the treatment difference is assessed against log-transformed equivalence limits. If the limits are \(L\) and \(U\), the two one-sided tests require evidence that the treatment difference exceeds \(L\) and is below \(U\). For a symmetric confidence-interval formulation, equivalence is concluded only when the appropriate confidence interval lies entirely inside \([L,U]\). Power is therefore affected by the standard error: greater uncertainty makes it harder for both one-sided tests to pass. The degrees of freedom also determine the critical value used in the test.

For a given allocation, exact power can be calculated from the noncentral \(t\) distribution under the planned treatment effect, standard error, degrees of freedom, and equivalence limits. A normal approximation can be useful for understanding allocation effects, but it is not a replacement for an exact calculation when final planning requires accurate power.

Planning Counts in Practice

For a fixed total, balance is the default when sequence-specific variability and costs are alike. If the planned total is odd, assigning one extra participant to either sequence gives the closest possible balance. Randomization should be arranged so that sequence counts remain close throughout recruitment, not merely at the end. Permuted blocks can help maintain that balance, while preserving random assignment within blocks.

The design target should usually be the evaluable sequence counts, not just the number randomized. If completion differs by sequence, equal enrollment can lead to unequal retained counts. For example, anticipated sequence-specific retention rates can be used to project retained counts and adjust recruitment targets. Those rates are uncertain, however; a planning adjustment should be based on credible operational information and should not be treated as a substitute for later methods addressing dropout.

Equal allocation is not automatically optimal when the two sequence-specific difference variances genuinely differ. If those variances are \(\sigma_{D,TR}^2\) and \(\sigma_{D,RT}^2\), then the variance of the treatment estimate is:

For a fixed total and equal cost per participant, minimizing this expression gives \(n_{TR}/n_{RT}=\sigma_{D,TR}/\sigma_{D,RT}\). This unequal-variance allocation is a specialized alternative, not a reason to assume that sequence variances differ. In the usual crossover model, the same within-participant variance applies to both sequences, so equal allocation remains optimal.

Diagnostics, Alternatives, and Common Pitfalls

Before analysis, report randomized and evaluable counts by sequence, and check that each sequence contains participants with usable outcomes in both periods. Review whether exclusions, missing periods, or protocol deviations are concentrated in one sequence. Examine the fitted model and residual variability for influential observations or marked sequence-specific differences. A large imbalance does not itself establish a model failure, but it can make conclusions more sensitive to a small number of participants in the smaller sequence.

The simple contrast of within-participant differences is a useful way to understand allocation in the two-sequence design. A linear model or mixed model can express the treatment and period effects directly and can accommodate some extensions, but its standard errors and degrees of freedom depend on the specified covariance structure and inferential method. The allocation principle remains: under comparable sequence variances, balancing the groups provides the most information for the treatment contrast. More complex crossover layouts require their own allocation and variance calculations; do not transfer the two-sequence formula without checking the design.

• Mistake: “Unequal counts bias the treatment effect.” Under the stated randomized model, imbalance increases variance but does not itself bias the contrast. Bias concerns arise from design or analysis problems such as differential carryover, informative missingness, or a misspecified model.
• Mistake: “A 24:16 allocation has fewer residual degrees of freedom than 20:20.” With the same complete total and the classical pooled analysis, both have \(N-2\) residual degrees of freedom. Their standard errors differ.
• Mistake: “Any allocation is acceptable if the total sample size is unchanged.” The variance contains \(1/n_{TR}+1/n_{RT}\); as one count becomes small, this term increases sharply.
• Mistake: “An observed imbalance should be fixed by changing assignments after enrollment.” Sequence allocation is set by the randomization plan. Investigators should not reassign participants based on observed outcomes; such changes can compromise randomization and inference.

The practical takeaway is to plan approximately equal sequence counts, maintain them with an appropriate randomization scheme, and calculate power using the actual intended counts. Treat degrees of freedom, precision, and power as related but distinct: in the basic complete-data model, imbalance changes precision and power while leaving \(N-2\) unchanged.