Randomization Schemes: How to Choose Before You Power

The sample size calculator takes the allocation ratio as an input (defaulting to 1:1) and assumes the analysis stratifies on whatever you stratified the randomization on. Setting the allocation ratio incorrectly — or ignoring stratification at analysis — makes the calculator's answer the wrong answer for the trial you actually run.

The randomization scheme determines three things that flow back into the calculator and the analysis:

• •The variance of the treatment-effect estimate. Blocking and stratification reduce variance when the stratification factors are prognostic. The standard sample size formula assumes simple randomization and over-estimates N when you stratify on factors that explain meaningful outcome variation.
• •The validity of standard test statistics. Conventional t-tests and chi-square tests assume independent observations within arms. Constrained randomization (minimization, biased coin) introduces dependence; ignoring it can inflate Type I error.
• •What you have to justify in the protocol. Regulators ask different questions for each scheme. Simple randomization is invisible; minimization triggers a paragraph in the FDA review. Pick the scheme that you can defend with the least friction.

Simple randomization

A coin flip (or its multi-arm equivalent) at each enrollment. Each patient is independently assigned to a treatment arm with the planned probability. Cheap, transparent, and statistically clean — conventional t-tests and chi-square tests apply without adjustment.

Permuted block randomization

Patients are randomized in fixed-size blocks (e.g., blocks of 4 or 6) that guarantee balance within each block. After every complete block, the allocation is exactly the planned ratio. The default for most modern clinical trials.

Block size matters: too small (e.g., 2) makes allocations predictable to enrolling investigators; too large defeats the balance benefit. A common choice for 1:1 allocation is variable block sizes (4, 6, or 8 mixed randomly) to blind the schedule.

Stratified randomization

Permuted blocks within strata defined by prognostic factors (e.g., disease severity, prior therapy, site). Each stratum gets its own block schedule; the result is balance on both treatment assignment and the stratification variables.

The trial must be analyzed accounting for the stratification — either by including stratification variables as covariates in the primary analysis (preferred) or by doing a stratified test. The sample size calculator's answer is usually slightly conservative when you stratify on prognostic factors that you also adjust for in analysis.

Minimization (covariate-adaptive randomization)

Dynamic balancing across many prognostic factors. Each patient is assigned to whichever arm minimizes overall imbalance on the covariates, with a small random component to preserve unpredictability. The most common implementation is the Pocock-Simon algorithm.

Use minimization when stratified randomization runs out of statistical headroom — typically when 4+ prognostic factors need balancing but stratifying on all of them produces empty strata. The FDA accepts minimization but the 2019 Adaptive Designs guidance requires that the inference scheme used be justified (often randomization-based tests or robust standard errors).

Quick reference: scheme vs. trial size

Central randomization and allocation concealment

Predictability discussed above is half the story; concealment is the other half. Even a well-chosen scheme leaks selection bias if enrolling investigators can see (or guess) the next assignment.

• •Central randomization (IWRS, IVRS, web-based) keeps the allocation sequence out of site hands. The site enrolls a patient, the central system returns the assignment.
• •Sealed sequentially-numbered envelopes are the low-tech alternative. They work only when the tamper-resistance is real (opaque, sequentially numbered, audited) — the default for sites without IWRS access.
• •Schedule access discipline. The full randomization schedule should live with the independent statistics group, not on a site server. Block sizes, stratification factors, and the seed are all bias-relevant information.

Stratification is not covariate adjustment

They pair well but neither replaces the other:

• •Stratification is a design-stage tool. It forces allocation balance on pre-randomization factors.
• •Covariate adjustment is an analysis-stage tool. It improves precision on the treatment-effect estimate and corrects residual imbalance, whether the trial stratified or not.

EMA's baseline-covariate guideline (EMA/CHMP/295050/2013) recommends including stratification factors in the primary analysis model when prespecified. The two tools combine additively — do both, but treat them as separate decisions.

The 50:50 baseline

For comparing two arms on a continuous, binary, or survival endpoint, the variance of the treatment-effect estimate is minimized when allocation is equal. This is exact for continuous outcomes with equal variances; approximate for binary and survival outcomes when the per-arm rates are close to each other. For most settings, 1:1 is statistically optimal.

The cost of unequal allocation

Moving away from 1:1 inflates total N. The penalty for a 2:1 allocation is $(1+r)^2 / (4r) = 9/8 \approx 1.125$ — you need about 12% more total patients to achieve the same power. The penalty grows quickly:

Two legitimate reasons to deviate

• •Ethics. When one arm carries a known burden (placebo in a serious disease, an invasive procedure), allocating fewer patients to it minimizes exposure. The cost in N is the price of the ethical posture.
• •Recruitment feasibility. When patients (or referring physicians) are unwilling to enroll in a trial with a high probability of placebo assignment, a 2:1 ratio toward active treatment can rescue the recruitment timeline. The biostatistician should make sure the sponsor accepts the N penalty before committing.

Response-adaptive randomization adjusts the allocation ratio during the trial based on accumulating outcome data: as evidence builds that one arm is performing better, new patients are preferentially allocated to it.

Two main families

• •Doubly-adaptive biased coin design (DBCD). Frequentist. Each new patient is assigned to a treatment with a probability that drives the cumulative allocation toward a target ratio that depends on the observed response rates (Hu & Zhang 2004).
• •Thompson sampling. Bayesian. Allocation probabilities are the posterior probabilities that each arm is the best. As the posterior concentrates on the better arm, allocation does too.

When RAR is worth the complexity

• •Multi-arm Phase II trials where the ethical priority is to minimize exposure to inferior arms.
• •Settings where outcomes accrue quickly enough that interim adjustments based on observed responses are operationally feasible (response rate measurable within weeks, not years).
• •Platform trials (REMAP-CAP-style) where multiple experimental arms compete against a shared control.

When it isn't

Most two-arm confirmatory trials. RAR introduces three complications that don't pay back at scale: (1) the time from outcome accrual to randomization adjustment must be short, which excludes most overall-survival endpoints; (2) Type I error control requires explicit simulation and pre-specification of the adaptation rule; (3) operational risk — an early run of bad luck on the new arm can lock the trial into allocations that take far longer to escape than the time savings RAR promised.

Allocation floors are the most important practical safeguard

Without a floor, an early run of poor outcomes on one arm drives that arm's allocation probability toward zero. Subsequent patients can't correct the noise because they never reach the arm. The standard fix is an allocation floor (commonly $\pi_k \geq 0.10$ or $0.15$ per arm) plus a burn-in period with fixed equal allocation (e.g., the first 20–40 patients per arm) before the adaptation begins. Without both safeguards, RAR can produce final estimates that look adaptive but are dominated by early-stage noise.

The randomization scheme determines which Zetyra calculator you should be running and how to set up its analysis plan.

Stratification and the analysis stage

Stratified randomization without stratified analysis is a half-measure: you spent the operational complexity to balance on a factor, then ignored it at analysis. The conventional recommendation (and FDA preference) is to include the stratification factors as covariates in the primary analysis model. This combines naturally with CUPED for further variance reduction on continuous endpoints.

But don't overfit. Including every stratification factor plus its interactions in a small trial produces a sparse, unstable model. EMA's baseline-covariate guideline (EMA/CHMP/295050/2013) recommends pre-specifying a small set of prognostic factors with clear scientific or clinical justification — not a kitchen-sink covariate list. If the cell counts within a stratum are very small (e.g., ≤ 5 per arm), consider collapsing levels or treating the factor as a covariate only at analysis without stratifying randomization on it.

• •

  Stratifying on too many factors

  Each additional stratification factor multiplies the number of strata. With 4 factors at 2 levels each, you have 16 strata; a 200-patient trial averages 12 patients per stratum, but the smallest strata may have 2–3. Empty strata break the balance guarantee. If you need to balance on more than 3 factors, move to minimization.

• •

  Block size of 4 in an open-label two-arm trial

  After observing 3 of the first 4 allocations, the 4th is deterministic. In an open-label trial, the enrolling investigator can predict it and may select patients to fit the expected arm — the textbook source of selection bias. Use variable block sizes (mix of 4, 6, 8) or larger fixed blocks in open-label settings.

• •

  Using RAR in a two-arm confirmatory trial where 50:50 was fine

  RAR's ethical benefit shines in multi-arm settings where many patients would otherwise see an inferior treatment. For a two-arm trial with a defined active control, 50:50 is statistically efficient and operationally simpler. RAR here trades clarity for marginal gains.

• •

  Powering for 1:1, then running 2:1 because the sponsor wanted it

  The calculator's N is wrong for the trial you actually ran. A 2:1 trial powered as 1:1 is roughly 12% under-powered. Set the allocation ratio in the calculator first; if the sponsor wants to change it later, re-run the calculation and increase N.

• •

  Treating minimization as a substitute for randomization

  Pure deterministic minimization is not random. Always include a stochastic component (typically 70–80% probability of the “minimizing” assignment) and document it. Otherwise the conventional inference framework breaks and you owe the regulator a randomization-based analysis.

• •Single-Arm Trials Guide — when no scheme is the right answer.
• •CUPED Guide — variance reduction via covariate adjustment at analysis. Interacts with stratification decisions.
• •Adaptive Randomization Technical Reference — full methodology for DBCD, Thompson sampling, and Pocock-Simon minimization.
• •Cluster Randomized Trials — when the randomization unit is a cluster (clinic, school, region) rather than an individual.

1. Pocock SJ, Simon R. Sequential treatment assignment with balancing for prognostic factors in the controlled clinical trial. Biometrics. 1975;31(1):103–115.
2. Efron B. Forcing a sequential experiment to be balanced. Biometrika. 1971;58(3):403–417.
3. Hu F, Zhang LX. Asymptotic properties of doubly adaptive biased coin designs for multitreatment clinical trials. The Annals of Statistics. 2004;32(1):268–301.
4. Thompson WR. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika. 1933;25(3–4): 285–294.
5. Berry SM, Carlin BP, Lee JJ, Müller P. Bayesian Adaptive Methods for Clinical Trials. CRC Press; 2010.
6. Senn S. Seven myths of randomisation in clinical trials. Statistics in Medicine. 2013;32(9):1439–1450.
7. Rosenberger WF, Lachin JM. Randomization in Clinical Trials: Theory and Practice. 2nd ed. Wiley; 2016.
8. U.S. Food and Drug Administration. Adaptive Designs for Clinical Trials of Drugs and Biologics: Guidance for Industry. November 2019.
9. EMA / CHMP. Guideline on Adjustment for Baseline Covariates in Clinical Trials. EMA/CHMP/295050/2013. February 2015.

Last updated: May 2026