Allocation, Minimization & Re-randomization — Reference Notes

Working notes built around a concrete case: 3 arms, 15 subjects/arm (N=45), multicenter, stratifying on wound size.
The thread runs from classical block randomization to why that case forces minimization, the EMA position on it, and the
re-randomization test that makes it defensible. Review-side flags are marked ⚑.

1. The allocation problem

Randomization exists to make treatment groups comparable on everything — measured and unmeasured — so the treatment
contrast isn’t confounded. Two failure modes it must guard against:
• Numeric imbalance — unequal arm sizes (costs power; worst at the optimal allocation ratio).
• Prognostic imbalance — arms differ on a covariate that drives outcome (costs validity; biases the unadjusted estimate).
Simple (complete) randomization — each patient an independent coin flip — controls neither during recruitment. Severe im-
balance is unlikely in a large trial but not impossible, and it’s a live risk at small N. Every method below is a constraint layered
on simple randomization to buy balance, and every constraint trades away some unpredictability to get it. That trade is the
through-line of these notes.

2. Block randomization
Origin. Fisher at Rothamsted, controlling a spatial fertility gradient across a field by grouping plots into blocks of similar fertility
and randomizing treatments within each block. In a trial the analogue of the spatial gradient is time: patient mix drifts over
recruitment (referral patterns, severity, centre case-load). Treat each run of consecutively-presenting patients as a block, balance
within it, and any time trend is absorbed.
Mechanics (block of 6, two arms V/P). Each block gets exactly 3 V + 3 P. Number of distinct balanced orderings:

6 6!
( )= = 20 (e.g. VVPPPV, VPVVPP … )
3 3! 3!

The 6! counts all orderings; the two 3! divide out the interchangeable V’s and P’s. Draw one of the 20 uniformly per block. The
all-V string VVVVVV simply isn’t in the urn — blocking deletes every imbalanced outcome from the sample space before
the draw.
The conditional-structure insight (this is the real content).
• Marginally, each patient still has P(V)= 12 — identical to a coin flip.
• Jointly, within-block assignments are negatively correlated. Once some are revealed, the rest tilt to restore the count.
After V,V,P,V in a block of 6, the last two are forced to P (conditional P(V)=0).
• That self-correction is the balance mechanism — imbalance can never exceed the block size.
The predictability cost is the same fact seen from the other side. The last allocation in a block is often deterministic; with
a fixed, known block size and any leakage of prior assignments, an investigator can anticipate forced slots → selection bias.
Mitigation: variable/random block sizes, concealed size, central allocation.
Generalization. Block size n, two arms: C(n, n/2) sequences. n=4 → 6; n=6 → 20; n=8 → 70. Larger blocks = less predictable
but looser balance between block boundaries.
⚑ Single small fixed block size + weak concealment = predictability hole. A stated total not divisible by the stated block size =
the description is incomplete or the final block was handled ad hoc — ask.

3. More than two arms, and the divisibility wrinkle

For k arms and a balanced block of size n (n/k per arm), distinct orderings = the multinomial coefficient:

1
 𝑛!
( 𝑛𝑘 )! 𝑘

Three arms:

Block size Per arm Distinct sequences Blocks to reach 45

3
3 1 3!/(1! ) = 6 15
6 2 6!/(2!3 ) = 90 7.5 ✗
9 3 9!/(3!3 ) = 1680 5
15 5 15!/(5!3 ) = 756,756 3

Divisibility. 45 = 3·15; block-compatible divisors are 3, 9, 15, 45. Block-of-6 doesn’t tile 45 (7.5 blocks) — can’t give exact
15/15/15 from 6s alone.
The fact that saves you: every balanced block contributes equally to all arms by construction. So any mix of balanced blocks
summing to 45 lands exactly on 15/15/15. E.g. 9+6+3+9+6+3+9 = 45 → each arm = 15. This is how variable-block lists are
built: mix sizes for concealment, per-arm balance falls out for free.
Three-arm predictability is worse for small blocks. In a block of 3 (one per arm) the third slot is fully determined once two
are seen — every third patient guessable. Block-of-3 is essentially unusable unblinded. Block-of-9 is the workhorse, often
mixed with 6 and 3.

4. Stratified randomization, and where it breaks

Stratified = a separate block-randomization list per stratum, where strata = the cross-classification of prognostic factors.
Guarantees balance within each completed block within each stratum.
The arithmetic that kills it at small N. Each stratum carries one possibly-incomplete final block; residual imbalances are
per-stratum and do not cancel. You need subjects/stratum ≳ 4× block size. For N=45:

Stratification # strata n/stratum Verdict

None / single site 1 45 tightest; ~guaranteed
15/15/15
One binary factor 2 22.5 fine
Centre (3) only 3 15 borderline with block 6
Centre (3) × sex (2) 6 7.5 fragments — most strata
never finish a block
+ age band 12+ <4 balance is fiction

Consequence people miss: once you stratify, exact 15/15/15 is no longer guaranteed — it becomes approximate. The exact
guarantee (blocks tiling 45) is traded away the moment you split. If exact equality matters and you have ≤ 1 site, don’t stratify
— run one variable-block list over 45.
For the case trial (multicenter + wound size = two factors): cross-classified strata = sites × wound-bands. Even the friendliest
case (2 bands × 3 sites = 6 strata, 7.5 each) is already too fragmented, and real trials have more sites. → stratified blocks can’t
carry two factors at N=45. This is the minimization trigger.

5. Margins vs cells — the conceptual core

The whole distinction is margins versus cells of the factor grid (rows = sites, columns = wound bands).
• Cross-classifying (stratified blocks) balances every cell — “Site 2 ∧ large wound” is its own mini-trial demanding
A=B=C inside it. Strong, but cuts N into many tiny piles that can’t fill blocks.

2
 • Marginal balance (minimization) balances the row totals and column totals separately and ignores the cells. Each
arm gets ≈ 13 of each site’s patients and ≈ 13 of each wound band’s patients — but Site-2-large might be 2-1-1, and that’s
allowed.
Why dropping the cell requirement rescues you: you’re no longer asking 45 patients to balance 6–12 things at once, only two
short lists (a 3-entry site margin, a 2-entry wound margin). Every patient contributes to one row total and one column total no
matter what, so the sample never fragments.
Why it’s the right trade: your analysis model adjusts for the main effects of site and wound size — not their full interaction.
You balance exactly what the model uses. Spending scarce patients to balance interaction cells you can’t even estimate at N=45
would be the wrong trade.
Card-deal intuition: making each player’s hand even on suit and even on color separately is easy; demanding every
specific suit-and-color combination be even is impossible with few cards. Balance the two things the analysis
examines; ignore the combinations nobody estimates.

6. Minimization (Pocock–Simon)
Algorithm, per arriving patient:
1. For each candidate arm, compute the marginal imbalance that assignment would create — summed over each factor’s
margin separately (here: site total + wound-band total). Imbalance metric = range (max−min) across arms, or variance.
2. Identify the arm(s) minimizing total imbalance.
3. Assign to a minimizer with probability p (the random element), else assign randomly. p ≈ 0.75–0.80 typical.
4. Commit; update the margin tallies.
It never consults the cell (intersection) counts — only the margins. That is marginal balance in code.
The random element is non-negotiable. Deterministic minimization (p=1) makes assignments predictable from history →
selection bias, and is the form regulators object to. The random tilt restores unpredictability and underpins a defensible inferential
basis. Cost: with a random element the grand total can wobble by ±1–2 (the case sim landed 15/14/16 in one seed, 15/15/15 in
another). A protocol claiming dynamic allocation and perfectly equal arms is mildly self-contradictory.
Continuous factors must be banded for minimization (Pocock–Simon can’t ingest cm2 raw). Pre-specify clinical cut-points in
the protocol. But adjust on the continuous value in the analysis — banding is a randomization convenience; discarding the
continuous information costs power you can’t spare at N=45.
Three arms: extends directly — compute total marginal imbalance for each of the three candidate assignments, bias toward the
minimizer.

7. The case trial, worked

Design: N=45, 3 arms, minimize on site (S1/S2/S3) + wound band (small <5 cm2 / large ≥ 5 cm2 ), p=0.80, central IWRS.
Arrival deliberately uneven (S1 recruits ~2× S3).
Result — the margins (what we protected) come out near-perfect despite uneven arrival:
Site margin — | A | B | C |, e.g. S1: 8/7/8, S2: 5/5/6, S3: 2/2/2. Wound margin — small: 9/8/9, large: 6/6/7.
Both factors within one patient of even across all three arms. Cells we ignored can be ragged (e.g. S3/large = 1/0/0) — by
design, and harmless because nothing estimates that cell.
Translation to protocol:
1. Pre-specify wound-size cut-points (keep band count minimal — more bands = finer margin = slightly looser balance).
2. Minimize centrally on site + band, p≈ 0.8.
3. Analyze with site + wound size (continuous) in the pre-specified primary model.
4. Report the random element, cut-points, and a re-randomization sensitivity check.

3
8. EMA position on dynamic allocation
The “strongly discouraged” line is from the superseded document. The much-quoted “Dynamic allocation is strongly dis-
couraged… without adequate and appropriate supporting/sensitivity analysis, an application is unlikely to be successful” is
from the old CPMP Points to Consider (CPMP/EWP/2863/99, 2003/04). That was replaced by the 2015 Guideline on ad-
justment for baseline covariates (EMA/CHMP/295050/2013), in force 1 Sep 2015. A protocol or reviewer quoting “strongly
discouraged” is citing retired guidance.
Current (2015) position = conditional acceptance, not prohibition. Requirements, all of which map onto the checklist above:
1. Pre-specification of the covariates in the primary analysis (prevents post-hoc covariate fishing, esp. in non-linear models).
2. Randomization factors must be in the analysis model — any factor used to balance allocation is adjusted for. (The
single most common inconsistency; a major-objection point.)
3. A random component is expected — the deterministic form is what drew the original hostility.
4. Sensitivity / robustness analysis showing the result doesn’t hinge on the allocation scheme.
The real inferential basis (Senn et al.). The concern was never bureaucratic: under dynamic allocation, allocation probabilities
depend on accumulating data, so the independence assumptions under a standard model-based test don’t automatically hold.
Resolution: adjust for the balancing covariates (restores test validity) and/or use a re-randomization test (conditions on the
true mechanism). No full statistical consensus exists, but this package is accepted practice.
⚑ The trap configuration (deserves a major objection): deterministic minimization, no random element, factors absent from
the analysis, no sensitivity check.
⚑ Currency: verify nothing has superseded EMA/CHMP/295050/2013 before citing it formally — the estimands / complex-
designs reflection-paper pipeline has been active.

9. Re-randomization test
Logic. Under the sharp null (no effect for any patient), each patient’s outcome is fixed regardless of arm. The only randomness
in the trial was the allocation. So build the null distribution of any test statistic by holding outcomes fixed and re-running the
randomness that produced the assignments, recomputing the statistic each time. The p-value = fraction of re-runs at least as
extreme as observed. No normality, no independence, no large-sample appeal — the reference distribution is constructed, not
assumed.
Why “condition on the actual mechanism” is load-bearing.
• A naïve permutation test shuffles labels freely — valid only under simple randomization, where all allocations were
equiprobable.
• Minimization allocations are sequential, dependent, non-uniform — wildly imbalanced allocations essentially never occur.
A free shuffle samples the wrong reference set.
• The re-randomization test generates each reference allocation by re-executing the actual minimization (same arrival
order, same covariates, same p). It samples the true design distribution.
• Permutation test = exchangeable shuffles; re-randomization test = re-run the mechanism. For minimization only the latter
is correct.
Procedure.
1. Compute observed statistic T_obs — ideally the covariate-adjusted effect from the pre-specified model (validity + effi-
ciency together).
2. Freeze outcomes, covariates, arrival order.
3. Re-run the minimization to draw A* (reproduce the random element).
4. Re-fit the model under A, compute T.
5. Repeat B times (Monte Carlo, ≥ 10,000; full enumeration infeasible for sequential designs).
6. Two-sided p = (1 + #{|T*| ≥ |T_obs|}) / (B+1). The “+1” includes the observed allocation → never anti-conservative.
Subtleties.
• Exact for the sharp null (no effect for anyone); your model targets the weak null (equal means). Pairing the re-
randomization test with the adjusted statistic makes it behave well for the weak null too. Reported as a robustness
companion to the model estimate, not a replacement.

4
 • Arrival order is part of the conditioning — fix it, re-randomize only the algorithm’s decisions.
• Re-fit the adjustment under each A* (don’t freeze it).
• Three arms → omnibus (model-based 𝜒2 /F across arms); pairwise contrasts legitimate with multiplicity control.
Case-trial result (N=45, real effects B+7 / C+12, adjusted F for arm).

Method p-value
Parametric F 0.0015
Re-randomization (correct) 0.0029
Naïve free-permutation 0.0017

All agree the effect is real — the usual outcome when the signal is clear. The instructive part is the null shapes: the re-
randomization null is more dispersed (95th pct ≈ 3.84) than the free-permutation null (95th pct ≈ 3.33). The naïve null is too
tight → anti-conservative here; in a borderline case (observed F ≈ 3.5) it would falsely reject where the correct test would not.
The nulls differ because you adjusted for the same covariates minimization balanced on; only the re-randomization test captures
the constrained design’s true null. Single-dataset p-values are illustrative; the structural finding (nulls differ, re-rand is the valid
reference) is the durable point.

10. Defensible package (the whole chain) & review checklist

To specify / expect in a protocol:
1. Minimize on the prognostic factors + site, with a random element (state p).
2. Pre-specify factors and any continuous-variable cut-points before unblinding.
3. Adjust for the randomization factors in the primary model (continuous covariates kept continuous).
4. Report a re-randomization test re-executing the exact algorithm, outcomes fixed — primary or co-primary inference.
Agreement with the model-based p demonstrates robustness; disagreement → trust the re-randomization test.
⚑ Reviewer flags:
• Randomization factors not in the analysis model → major objection (EMA 2015 §2).
• Deterministic minimization / no random element → predictability + validity hole.
• Continuous covariate banded for randomization and analyzed only as bands → power thrown away; ask for continuous
adjustment.
• No sensitivity/re-randomization analysis → missing the EMA-expected support.
• Naïve permutation or unadjusted parametric test offered as the validity argument under minimization → wrong reference
set.
• Stated arm total not divisible by stated block size, or single small fixed block → incomplete description / concealment
concern.
• “Dynamic allocation strongly discouraged” cited as current EMA → it’s from the retired CPMP PtC.

Formula quick-reference

Quantity Formula
Balanced block sequences, 2 arms C(n, n/2)
Balanced block sequences, k arms n! / ((n/k)!)^k
Subjects/stratum rule of thumb ≳ 4 × block size
Re-randomization p (two-sided) (1 + #{|𝑇 ∗ | ≥ |𝑇obs |})/(𝐵 + 1)
Block-compatible divisors of 45 (3 arms) 3, 9, 15, 45

Key refs: ICH E9 / E9(R1); EMA/CHMP/295050/2013 (2015); Pocock & Simon (1975); Taves (1974); Senn, Statistical Issues
in Drug Development.