Predictability From Fixed Block Sizes

Why a Randomized Schedule Can Become Predictable

Permuted blocks are used to keep the numbers assigned to treatment arms close as enrollment proceeds. But balance has a consequence: once some assignments in a block are known, the assignments still available are constrained by the block’s fixed composition. The next treatment may therefore be less than equally likely to be either arm, and at the end of a block it can be determined with certainty.

This is not evidence that the schedule was generated incorrectly. It is a consequence of sampling without replacement from a known set of treatment slots. The practical concern is whether someone involved in enrollment can use this information to anticipate an assignment before deciding which eligible participant to enroll.

Conditional Probabilities Within a Block

Consider a block containing \(a\) assignments to arm A and \(b\) assignments to arm B, so its size is \(m=a+b\). Assume that all permutations of these \(m\) assignments are equally likely. Suppose the block boundary and composition are known, and the assignments already made in that block have been observed accurately. If \(r_A\) A assignments and \(r_B\) B assignments remain, then \(r_A+r_B\) slots remain. By symmetry among those remaining slots, the next assignment has probabilities

These are conditional probabilities: they apply after taking account of the known assignments in the current block. They are not necessarily \(1/2\), even in a block that contains equal numbers of A and B assignments. If only B slots remain, then \(r_A=0\), and the next assignment must be B.

The calculation depends on specific conditions. The observer must know which block is in progress, know the treatment counts assigned to that block, and know the assignments already used. The assignments must have been permuted uniformly, and there must be no unrecorded deviations from the schedule. If the block boundary is uncertain or the composition is concealed, this direct calculation is not available without additional information; uncertainty about the boundary does not, by itself, guarantee that prediction is impossible.

A Short Justification

Among the \(r_A+r_B\) remaining positions, each position is equally likely to be next under a uniform random permutation. Exactly \(r_A\) of those positions carry A. Equivalently, the remaining multiset can be ordered in \((r_A+r_B)!\) ways, while the orders with A first number \(r_A(r_A+r_B-1)!\). Therefore,

The same argument gives the probability for B. The formula is about the next position in the block, not about the probability that a particular patient benefits from treatment or about the probability that an allocation method is ethically acceptable.

Why Predictability Matters During Enrollment

The statistical calculation becomes a conduct concern when someone can use a likely or known assignment to influence enrollment. For example, a person who can see the allocation history might delay enrolling an eligible patient, accelerate enrollment, or affect when eligibility is confirmed in the hope of obtaining a preferred arm for that patient. Even if the randomization schedule was generated correctly, using assignment foreknowledge to shape who enters at a particular point can undermine the intended random allocation.

The relevant protection is not merely that the original list was random. The allocation needs to remain concealed from people who could alter enrollment or eligibility decisions until the participant has been irrevocably entered and the assignment is issued. In practice, this supports procedures such as restricted access to the schedule and a secure allocation service that reveals an assignment only when an eligible participant is enrolled. The appropriate safeguards depend on who can see the schedule, who controls enrollment, and how the allocation system operates.

Predictability and concealment are related but distinct. Predictability describes what can be inferred from available information. Concealment describes whether the people making enrollment decisions can access that information before enrollment. A schedule can have mathematically predictable late-block assignments while strong access controls prevent recruiters from knowing them. Conversely, an assignment process can be vulnerable even when no one can calculate the next assignment with certainty, if people can exploit a substantial shift in its probabilities.

Diagnostics, Alternatives, and Limits

A useful diagnostic is to reconstruct the current block’s remaining counts after each assignment and calculate the conditional probabilities. If the sequence is being monitored operationally, the review should check whether the block size and boundary are visible, who can access the assignment history, and whether any assignment was skipped, delayed, or otherwise deviated from the schedule. A calculation based on incorrect counts or an assumed block boundary can give a false impression of certainty.

Fixed permuted blocks trade unrestricted allocation for local balance. Complete randomization avoids fixed local block boundaries but can allow greater chance imbalance during enrollment. Randomizing the block size can make it harder for an observer to know which block position is current, although it does not make every allocation unpredictable in every circumstance. The central comparison is therefore not “predictable versus random”: each approach is randomized under its design, but the information revealed by past assignments differs.

Fixed blocks also do not guarantee equal allocation at every point in the trial. They guarantee the specified counts at the end of each completed block, assuming the schedule is followed. During a block, the running totals can differ. Nor does arithmetic alone quantify the practical risk of enrollment manipulation: that assessment also depends on access, timing, staff roles, and compliance procedures.

Common Mistake

A common error is to reason, “The block was randomized, so the next assignment still has probability \(1/2\) for each arm.” Randomization gives probabilities before the relevant assignments are observed; conditioning on the observed history can change those probabilities. For a six-position block with three A and three B assignments, after observing A, A, B, one A and two B assignments remain, so the next probability of A is \(1/3\), not \(1/2\). The correction is to count the unused assignments in the current block and divide by the number of remaining positions.

A second error is to say that every fixed-block sequence is predictable from its first assignment. It is not. Some histories leave a balanced mix of arms and therefore do not identify the next assignment; certainty occurs only when all remaining slots belong to one arm. The correct statement is conditional: predictability changes as assignments are observed and depends on the remaining block composition.