Changing Randomization Ratios at the Transition

Why the Transition Ratio Matters

A seamless trial may retain a treatment at the phase boundary but change how new participants are allocated. For example, an early stage may use equal randomization, while the confirmatory stage assigns twice as many participants to treatment as to control. The treatment effect remains expressed on the same outcome scale, but its variance depends on the number assigned to each arm. Consequently, the two stages may contribute different amounts of information even if their sample sizes are similar.

This tutorial develops the variance calculation for a continuous outcome and explains how it informs stagewise combination testing. The distinction is important: changing an allocation ratio changes precision, but does not by itself invalidate a combination test. Validity depends on the stagewise tests, their dependence, and the combination rule—not on the ratios being equal.

Model and assumptions: In stage \(k\), let \(n_{Tk}\) and \(n_{Ck}\) be the numbers assigned to treatment and control. Assume independent participants, independent stage cohorts, and normally distributed outcomes with a common variance \(\sigma^2\) in both arms and stages. The stage-\(k\) treatment effect is the treatment mean minus the control mean, \(\hat{\theta}_k=\bar{x}_{Tk}-\bar{x}_{Ck}\). The variance calculations below use a known \(\sigma^2\); in practice it may be estimated, in which case the stagewise test must account for that estimation. The endpoint, direction of benefit, and treatment-effect estimand must be aligned across stages.

Variance Under a Stage-Specific Allocation Ratio

Because the two arm means are based on independent observations, the variance of their difference is the sum of their variances:

\operatorname{Var}(\hat{\theta}_k) =\operatorname{Var}(\bar{x}_{Tk})+\operatorname{Var}(\bar{x}_{Ck}) =\frac{\sigma^2}{n_{Tk}}+\frac{\sigma^2}{n_{Ck}}.

Write the treatment-to-control allocation ratio as \(r_k=n_{Tk}/n_{Ck}\), and the stage total as \(N_k=n_{Tk}+n_{Ck}\). Then \(n_{Tk}=N_k r_k/(1+r_k)\) and \(n_{Ck}=N_k/(1+r_k)\). Substitution gives:

\operatorname{Var}(\hat{\theta}_k) =\frac{\sigma^2}{N_k}\left(\frac{(1+r_k)^2}{r_k}\right), \qquad SE(\hat{\theta}_k) =\sigma\sqrt{\frac{(1+r_k)^2}{r_k N_k}}.

The allocation factor \((1+r_k)^2/r_k\) is minimized at \(r_k=1\) when the arm variances and per-participant costs are equal. This does not mean equal allocation is always the design choice: a trial may value additional treatment exposure, safety information, or practical constraints. It does mean that, at a fixed total sample size under this model, moving away from equal allocation costs precision.

Derivation: Since \((1+r)^2/r=r+2+1/r\), its derivative for \(r>0\) is \(1-1/r^2\). This derivative is negative for \(0<r<1\), zero at \(r=1\), and positive for \(r>1\). Thus the factor decreases up to equal allocation and increases afterward; its minimum is \(4\) at \(r=1\). The variance at that allocation is \(4\sigma^2/N\).

Worked Example: Comparing Ratios at a Fixed Stage Size

Suppose a hypothetical stage has \(N=150\) participants and outcome standard deviation \(\sigma=8\) points. Compare equal allocation with a 2:1 treatment-to-control allocation. The model assumes the same outcome variance in both arms.

With equal allocation, \(n_T=n_C=75\), so:

\operatorname{Var}(\hat{\theta}) =8^2\left(\frac{1}{75}+\frac{1}{75}\right) =64\left(\frac{2}{75}\right) =1.7067\ \text{points}^2, \qquad SE=1.306\ \text{points}.

With ratio \(r=2\), the allocation is \(n_T=100\) and \(n_C=50\). Therefore:

\operatorname{Var}(\hat{\theta}) =64\left(\frac{1}{100}+\frac{1}{50}\right) =64(0.03000) =1.920\ \text{points}^2, \qquad SE=1.386\ \text{points}.

The 2:1 allocation has \(1.920/1.7067=1.125\), or 12.5%, greater variance for the treatment-effect estimate at the same total sample size. This comparison assumes equal arm variances and equal participant costs; it is not a general claim that equal allocation is best under every design objective.

Information and Combination Weights

For the known-variance model, define stage information as the inverse variance, \(I_k=1/\operatorname{Var}(\hat{\theta}_k)\). More information means a more precise stage estimate. The standardized statistic for a one-sided test of no benefit at the boundary is \(Z_k=\hat{\theta}_k/SE(\hat{\theta}_k)\), which has a standard normal distribution under that boundary null when the model holds.

When stage cohorts are independent, their stage statistics are independent under the null, provided each statistic is calculated from that stage’s data using its actual design and variance. A fixed-weight combination statistic can then be formed as \(Z_{\mathrm{comb}}=w_1Z_1+w_2Z_2\), with \(w_1^2+w_2^2=1\). If weights are chosen proportional to the square roots of the stages’ information, then \(w_k=\sqrt{I_k/(I_1+I_2)}\). This choice reflects relative precision; it is not required for the validity of every combination test.

Justification: Under the boundary null, suppose \(Z_1\) and \(Z_2\) are independent standard normal variables and the weights are fixed in advance. A linear combination of independent normal variables is normal, with mean zero and variance \(w_1^2+w_2^2=1\). Therefore \(Z_{\mathrm{comb}}\sim N(0,1)\) under that null. The one-sided test rejects for a sufficiently large positive value. If stagewise p-values rather than z-statistics are combined, the combination function must likewise be calibrated for valid stagewise p-values and its specified weights.

Worked Example: Information After a Ratio Change

Consider two hypothetical independent stages with the same outcome standard deviation, \(\sigma=10\) points. Stage 1 enrolls 120 participants equally: 60 per arm. Stage 2 enrolls 180 participants at a 2:1 ratio: 120 on treatment and 60 on control. Assume the arm sizes are as planned and the model’s common-variance assumption is reasonable.

The stage variances and information are:

V_1=100\left(\frac{1}{60}+\frac{1}{60}\right)=3.3333, \qquad I_1=\frac{1}{3.3333}=0.3000,

V_2=100\left(\frac{1}{120}+\frac{1}{60}\right)=2.500, \qquad I_2=\frac{1}{2.500}=0.4000.

Although stage 2 has 1.5 times as many participants, it has only \(0.4000/0.3000=1.333\) times the information. Its 2:1 allocation is less precise per participant than equal allocation would be under the stated assumptions. Information-based fixed weights are:

w_1=\sqrt{\frac{0.3000}{0.7000}}=\sqrt{\frac{3}{7}}=0.6547, \qquad w_2=\sqrt{\frac{0.4000}{0.7000}}=\sqrt{\frac{4}{7}}=0.7559.

The weights reflect stage information, not sample-size proportions alone. Using the stage totals to set weights while ignoring the different ratios would not match this information calculation.

Testing When the Ratio Is Changed at the Boundary

A ratio may be selected at the transition using earlier trial information. For a stagewise combination test to remain valid, the stage-2 test must be valid under the null conditional on the information used to choose the stage-2 design. In this setting, the new stage’s allocation should be implemented as specified, and its statistic or p-value must use the actual arm sizes and an appropriate variance model. A ratio change alone does not make the stage-2 p-value invalid.

A straightforward approach is to define a combination rule and its weights in advance, then combine valid stagewise evidence. Under the null, the stage-2 p-value should be conditionally valid given the stage-1 data and the chosen allocation. This conditional validity, together with the combination procedure’s assumptions, is what supports type I error control. For example, a standard-normal combination statistic with fixed weights requires independent, standard-normal stage statistics under the null; other combination functions require their own calibration.

There is an important distinction between adapting the allocation and adapting the combination weights. A ratio chosen at the transition changes \(V_2\) and hence the distributional calculation for stage 2. It does not automatically authorize choosing weights after observing stage-1 results and then treating those weights as fixed from the outset. Data-dependent weights can alter the null distribution of the combined statistic. Such adaptation needs a combination or conditional-error procedure that explicitly accounts for the rule. Merely inserting the observed ratio into an otherwise unadjusted formula is not a general validity argument.

Worked Example: A Fixed-Weight Combination Test

Use the two stages and information-based fixed weights from the previous example. Suppose the hypothetical observed stage-1 treatment effect is 0.00 points, giving \(Z_1=0.000\). Suppose the stage-2 effect estimate is \(4.886\) points. Its standard error is \(\sqrt{2.500}=1.581\) points, so \(Z_2\) is approximately \(4.886/1.581=3.090\). Take the one-sided null to be no positive treatment effect, and assume the outcome direction and estimand are aligned across stages.

The combined statistic is:

Z_{\mathrm{comb}} =(0.6547)(0.000)+(0.7559)(3.090) \approx 2.336.

The approximate upper-tail standard normal p-value is \(P(Z\geq 2.336)=0.009746\). At a one-sided significance level of \(\alpha=0.025\), this is below \(\alpha\), so the fixed-weight combination test rejects the boundary null. This conclusion relies on the independent-stage and correctly specified stagewise-test assumptions; it is not obtained by pooling the estimates while ignoring their variances.

Diagnostics, Alternatives, and Limitations

Before interpreting a variance or combination result, check that the inputs describe the trial actually conducted. Confirm the number analyzed in each arm at each stage, the allocation ratio achieved, and whether the stated variance model is plausible. The common-variance formula can be inappropriate if arm variances differ materially; with independent arms, the more general variance is \(\sigma_{Tk}^2/n_{Tk}+\sigma_{Ck}^2/n_{Ck}\). If the outcome is binary, time-to-event, or adjusted for baseline covariates, its information calculation may require a different model rather than this continuous-outcome formula.

Also check that stages truly provide independent evidence. Reusing participants or a shared control can induce dependence, so the independent-stage combination justification does not apply without adjustment. A change in measurement, population, or estimand can also make two stage effects poorly aligned; variance arithmetic cannot resolve that scientific problem. Finally, estimated variances and small samples may require t-based or other calibrated stagewise tests rather than treating each standardized statistic as exactly normal.

Approach	What it uses	Main consideration
Fixed-weight stagewise combination	Valid stage-specific statistics or p-values and a prespecified combination rule	Can accommodate different stage variances; validity depends on stagewise validity and the combination rule’s assumptions.
Information-weighted combination	Stage information, often through weights proportional to square roots of information	Gives more weight to more precise stages, but weights must be specified or adapted under a valid rule.
Pooled estimate and test	Stage estimates and their variances in a single analysis	Can be efficient under a correctly specified common-effect model, but post-transition selection or adaptation may require more than an ordinary pooled analysis.

A common mistake is to assume that a 2:1 ratio gives stage 2 twice the information of an equally sized 1:1 stage because there are twice as many treatment participants per control participant. The correction is to calculate the variance from both arm sizes: treatment and control observations both contribute to the precision of a difference. Another mistake is to use a stage’s planned ratio after actual arm sizes differ; the variance should reflect the analyzed data and the analysis model. Neither mistake is repaired simply by changing the combination weights after seeing the results.

Key takeaway: Calculate each stage’s treatment-effect variance using its own arm sizes. A changed ratio changes information and may change appropriate fixed weights, but valid stagewise combination testing rests on valid stage-specific evidence, appropriate dependence assumptions, and a calibrated combination rule.

Check Your Understanding

Answer each question using the stated assumptions. Brief solutions follow.

With a common standard deviation of 6 points, what is the treatment-effect variance for a stage with 40 participants per arm?
A stage has 120 participants and ratio \(r=2\). How many are assigned to each arm, and what is the allocation factor \((1+r)^2/r\)?
Why does a 2:1 ratio not generally double the information at a fixed total sample size?
Under what conditions does \(w_1Z_1+w_2Z_2\) have a standard normal null distribution?
Does selecting the stage-2 ratio at the transition automatically justify choosing combination weights after observing stage-1 results? Explain.

Solutions. (1) \(36(1/40+1/40)=1.800\) points squared. (2) \(n_T=80\), \(n_C=40\), and the factor is \(9/2=4.500\). (3) Both arm sizes determine the variance; treatment and control counts enter through their reciprocal contributions, and unequal allocation sacrifices precision under equal variances and costs. (4) The stage statistics must be independent standard normal variables under the null, with weights fixed and \(w_1^2+w_2^2=1\). (5) No. Data-dependent weights require a combination or conditional-error procedure that accounts for their selection; treating them as prespecified can give the wrong null distribution.